Top Banner
ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data Mining with Big Data BY DICKSON WAI HEI LAM A project submitted in partial fulfillment of the requirements for the degree of MASTER OF SCIENCE in INFORMATION SYSTEMS Athabasca, Alberta July, 2014 © Dickson Wai Hei Lam, 2014
161

ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

Mar 07, 2018

Download

Documents

vanngoc
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Page 1: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

ATHABASCA UNIVERSITY

A Survey of Predictive Analytics in Data Mining with Big Data

BY

DICKSON WAI HEI LAM

A project submitted in partial fulfillment

of the requirements for the degree of

MASTER OF SCIENCE in INFORMATION SYSTEMS

Athabasca, Alberta

July, 2014

© Dickson Wai Hei Lam, 2014

Page 2: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 2

2

DEDICATION

I dedicate this essay to my parents Lai Hoi Yin and Lam Bing Fat, my brother Lam Tung

Hay, my sister Tracy Lam and my brother-in-law Matthew Chao. Last but not least, my niece and

my two nephews, Kasey, Evan and Gabriel. For their understandings, supports and

encouragements throughout the entire MScIS program in the past four years.

I am especially thankful to my wife, Tracy Low, for her love, patience and support during

the past four years. It is with her encouragement that I found the confidence to have embarked on

this journey. It is also with her perseverance that we are able to conclude this journey with this

essay.

Page 3: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 3

3

ABSTRACT

This paper explores the area of Predictive Analytics in combination of Data Mining and

Big Data. The survey indicates an accelerated adoption in the aforementioned technologies in

recent years. Businesses and researchers alike take great interests in furthering the use of

Predictive Analytics in enhancing Business Intelligence and forecasting ability across a wide

range of applications. This research essay explained some of the underpinnings in enabling

predictive capabilities in data analysis and Data Mining. Also, it incorporated the characteristics

of Big Data as the supplementary enabler to augment the way we perceive Data Mining.

Predictive analytics is the next frontier for innovation that is built based on century old concepts

and techniques such as mathematical analysis and statistical analysis.

Keywords: Predictive Analytics, Data Mining, Big Data, Analytics, Statistical Analysis,

Machine Learning

Page 4: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 4

4

ACKNOWLEDGMENTS

I would like to express my deepest gratitude to the faculty members of Master of Science

in Information Systems (MScIS). I would also like to express my sincere appreciation to the

MScIS Graduate Program Director, Dr. Larbi Esmahi, for his acceptance of being my essay

supervisor. I would like to thank him for his advice and endless support throughout the

development of this essay.

Page 5: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 5

5

TABLE OF CONTENT

DEDICATION ............................................................................................................................................... 2

ABSTRACT ................................................................................................................................................... 3

ACKNOWLEDGMENTS ............................................................................................................................. 4

TABLE OF CONTENT ................................................................................................................................ 5

LIST OF TABLES AND FIGURES ............................................................................................................ 9

CHAPTER I ................................................................................................................................................. 12

INTRODUCTION .................................................................................................................................... 12

Statement of the Purpose ..................................................................................................................... 12

Research Problem ................................................................................................................................ 14

Background ..................................................................................................................................................... 14

The Problems .................................................................................................................................................. 16

The Definitions of Predictive Analytics and Big Data .................................................................................... 17

Organization of the Remaining Chapters ............................................................................................ 18

CHAPTER II ............................................................................................................................................... 20

REVIEW OF RELATED LITERATURE ................................................................................................ 20

Introduction ......................................................................................................................................... 20

Explanatory versus Predictive Modeling ............................................................................................. 23

The Role of Predictive Model ......................................................................................................................... 24

Underlying Models Differences ...................................................................................................................... 24

Data Mining and Big Data .................................................................................................................. 28

The Big Data Problems ................................................................................................................................... 29

The NoSQL Solution ....................................................................................................................................... 31

The Apache Hadoop Platform ......................................................................................................................... 33

NoSQL Data Model versus Relational Data Model ........................................................................................ 35

Page 6: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 6

6

Predictive Analytics Versus Other Forecasting and Mining Methods ................................................. 36

The Predictive Model Markup Language (PMML) ......................................................................................... 40

PMML Adoption ........................................................................................................................................ 40

PMML Document Structure ....................................................................................................................... 41

PMML Interoperability and Application .................................................................................................... 44

PMML Enabled Architecture ..................................................................................................................... 46

Cloud Computing ............................................................................................................................................ 48

Dichotomy of Predictive Analytics Skillsets ................................................................................................... 50

Predictive Analytics Performance Optimization ............................................................................................. 52

CHAPTER III .............................................................................................................................................. 57

PREDICTIVE ANALYTICS APPLICATIONS .................................................................................................... 57

Social Computing ................................................................................................................................. 58

Large-Scale Machine Learning at Twitter Inc. ................................................................................................ 58

Network Relationship ...................................................................................................................................... 62

Education Support ............................................................................................................................... 64

Predicting At-Risk Students ............................................................................................................................ 64

Predict Course Success.................................................................................................................................... 66

Video Gaming ...................................................................................................................................... 67

Law Enforcement ................................................................................................................................. 70

Business Applications .......................................................................................................................... 76

Financial Engineering ......................................................................................................................... 78

Summary .............................................................................................................................................. 81

CHAPTER IV .............................................................................................................................................. 84

METHODOLOGY ................................................................................................................................... 84

Preliminary Discussion ........................................................................................................................ 86

Preliminary Analysis of the Topic Data Using Google Trends............................................................ 88

Discussion ............................................................................................................................................ 90

Context-Awareness and Big Data ................................................................................................................... 90

Page 7: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 7

7

Basic Statistical Methods and Techniques ...................................................................................................... 97

Data Mining Methods and Techniques in Predictive Analytics .................................................................... 101

Classification ........................................................................................................................................... 101

Regression ................................................................................................................................................ 103

Clustering ................................................................................................................................................. 105

Artificial Neural Network ........................................................................................................................ 106

Conclusion ......................................................................................................................................... 109

CHAPTER V ............................................................................................................................................. 112

ISSUES, CHALLENGES, AND TRENDS ............................................................................................ 112

Introduction ....................................................................................................................................... 112

Big Data Issues, Challenges and Trends in Predictive Analytics ...................................................... 115

Trend in Big Data Application ........................................................................................................... 118

Predictive Analytics Issues, Challenges and Trends .......................................................................... 122

The Ensemble Approach ............................................................................................................................... 122

The Concept Drift ......................................................................................................................................... 124

Trends and Advancements ............................................................................................................................ 124

Ethical Concerns and Issues .......................................................................................................................... 128

Conclusion ......................................................................................................................................... 130

CHAPTER VI ............................................................................................................................................ 132

CONCLUSIONS AND RECOMMENDATIONS .................................................................................. 132

Conclusions........................................................................................................................................ 132

Suggestions for Further Research ..................................................................................................... 133

SOA .............................................................................................................................................................. 133

Real-time analytics ........................................................................................................................................ 134

NoSQL .......................................................................................................................................................... 135

REFERENCES .......................................................................................................................................... 137

APPENDIX A – PMML CODE ............................................................................................................... 155

PMML CODE EXAMPLE .......................................................................................................................... 155

Page 8: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 8

8

PMML CODE EXAMPLE - HEADER SECTION ........................................................................................... 155

PMML CODE EXAMPLE - DATADICTIONARY SECTION ........................................................................... 155

PMML CODE EXAMPLE - TRANSFORMATIONDICTIONARY SECTION ...................................................... 155

PMML CODE EXAMPLE - MODEL SECTION – SUPPORT VECTOR MACHINE ............................................ 156

APPENDIX B – RESEARCH TOOLS .................................................................................................... 159

PRODUCTIVITY SOFTWARE ...................................................................................................................... 159

INTERNET BROWSERS .............................................................................................................................. 159

OPEN SOURCE PREDICTIVE ANALYTICS AND DATA MINING TOOLS ....................................................... 159

PYTHON RELATED STATISTICAL LIBRARIES ............................................................................................ 159

LITERATURE SEARCH ENGINES ............................................................................................................... 160

RESEARCH PAPER ONLINE DATABASES .................................................................................................. 161

RESEARCH MANAGEMENT TOOLS AND SERVICES ................................................................................... 161

ONLINE COMMUNITIES ............................................................................................................................ 161

Page 9: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 9

9

LIST OF TABLES AND FIGURES

Figure 1: Business Intelligence Taxonomy ................................................................................... 16

Figure 2: Decision Support Process .............................................................................................. 21

Figure 3: Analytics Processing Taxonomy .................................................................................... 22

Figure 4: Explanatory Model ........................................................................................................ 25

Figure 5: Predictive Model ........................................................................................................... 25

Figure 6: Descriptive Analytics Taxonomy .................................................................................. 27

Figure 7: Data mining adopts techniques from many domains (Han, Kamber, & Pei, Data

Mining: Concepts and Techniques, Third Edition, 2011) ............................................................. 29

Figure 8: HDFS Architecture (HDFS Architecture Guide, 2014)................................................. 33

Figure 9: Predictive Analytics in Data Mining ............................................................................. 38

Figure 10: Predictive Analytics Taxonomy ................................................................................... 39

Figure 11: PMML Schema (src: http://www.dmg.org) ................................................................. 42

Figure 12: Life Cycle of Data Mining Project using ADAPA (Guazzelli, Stathatos, & Zeller,

2009) ............................................................................................................................................. 45

Figure 13: EMC Greenplum MPP database architecture (Das, Fratkin, Gorajek, Stathatos, &

Gajjar, 2011).................................................................................................................................. 47

Figure 14: Automated modeling workflow (Kridel & Dolk, 2013) .............................................. 52

Figure 15: General Forecasting Process (Fischer, et al., 2013) .................................................... 54

Figure 16: Extension of 3-layer schema architecture (Fischer, et al., 2013) ................................ 56

Figure 17: Bayesian Net Model of Intent to Proliferate (Sanfilippo, et al., 2011) ....................... 69

Page 10: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 10

10

Figure 18: Group Recidivism Rates (Jennings & M.C.J, 2006). Note: n=number of individual per

category (scale) and per risk class, %=percentage of the individual in a specific class were re-

arrested. ......................................................................................................................................... 72

Figure 19: Bayes' theorem formula ............................................................................................... 77

Figure 20: "Predictive Analytics" Search Term - Google Trends Chart, February 2014 .............. 88

Figure 21: "Data Mining" Search Term - Google Trends Chart, February 2014 .......................... 89

Figure 22: "Big Data" Search Term - Google Trends Chart, February 2014 ................................ 89

Figure 23: Combined Search Terms - Google Trends Chart, February 2014. Blue color =

Predictive Analytics, Red Color = Data Mining, Orange Color = Big Data ................................. 89

Figure 24: The evolution of context definition (Kiseleva, 2013) ................................................. 91

Figure 25: Layered conceptual framework for context-aware systems (Baldauf, Dustdar, &

Rosenberg, 2007) .......................................................................................................................... 92

Figure 26: Meta-Learning Architecture (Singh & Rao, 2013) ...................................................... 96

Figure 27: An example of context-aware system design (Kiseleva, 2013) ................................... 96

Figure 28: Context managing framework architecture (Baldauf, Dustdar, & Rosenberg, 2007) . 97

Figure 29: An example of a normal distribution "Bell Curve" ..................................................... 98

Figure 30: Variance formula ......................................................................................................... 98

Figure 31: Standard Deviation formula 1 ..................................................................................... 99

Figure 32: Standard Deviation formula 2 ..................................................................................... 99

Figure 33: Correlation Coefficient formula .................................................................................. 99

Figure 34: A Bayes net for the medical diagnosis example (Patterns of Inference, 2014) ......... 103

Figure 35: A visualized example of linear regression (Natural Resources Canada, 2012) ......... 104

Page 11: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 11

11

Figure 36: Clustering of a set of objects using the k-means method; for (b) update cluster centers

and reassign objects accordingly (the mean of each cluster is marked by a C) (Han, Kamber, &

Pei, Data Mining: Concepts and Techniques, Third Edition, 2011) ........................................... 106

Figure 37: Multi-layer feed forward neural network (Han, Kamber, & Pei, Data Mining:

Concepts and Techniques, Third Edition, 2011) ......................................................................... 108

Figure 38: Binary Threshold Neuron formula ............................................................................ 109

Figure 39: An example of binary response dependent variable (x=income, y=gender) ..............114

Figure 40: Value of knowledge about event (Fülöp, et al., 2012) ............................................... 120

Figure 41: CEP-PA conceptual framework ................................................................................. 121

Page 12: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 12

12

CHAPTER I

INTRODUCTION

Statement of the Purpose

The cumulating amount of data has pressured researchers and practitioners to devise new

techniques and data processing models to tap into the invaluable source of Big Data. One such

usage in extracting knowledge from the vast amount of data is in Predictive Analytics which

allows us to gain insights in predicting unknown events and future activities. Within the context

of Data Mining, Predictive Analytics pairs with statistical analysis to provide a very interesting

combination of techniques for knowledge discovery.

The predictive nature within the domain of statistical analysis paves the way for a wide

variety of real life applications. From clinical analytics in Clinical Decision Support System

(CDSS) to business analytics in Operations Research (OR), Predictive Analytics aids decision

makers to make choices and solve problems that have long lasting impacts. The foundation that

brings new opportunities and possibilities of continuous improvement came from the centuries

old disciplines of mathematics and statistics. From data collection to data analysis, mathematics

and statistics interweave into every aspect of scientific data research endeavors. Predictive

Analytics is fundamentally dependent on the disciplines of mathematics and statistics. In fact, the

roles that the two disciplines play are crucial to the evolutionary development process of Data

Mining and Big Data, which is still shaping our collective understandings of Predictive

Analytics.

The intent of the selected research topic is to survey the current landscape of Predictive

Analytics and Data Mining within the context of Big Data. Although many underlying

components of Data Mining and Big Data have been the subject of discussions and research

Page 13: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 13

13

interests for many years, the demand for better data management and data analysis tools continue

to accelerate. One of these trends in this regard is the rise of NoSQL databases which challenge

even the relational data model, a prevalent and dominant database model design since the second

half of twentieth century (Codd, 1970).

The purpose of this paper is to explore Predictive Analytics in conjunction with Data

Mining and Big Data. This research will also touch on the concerns related to cloud computing,

mobile computing and social computing to the extent that helps to solidify the discussion points

on the application of Predictive Analytics. During the initial phrase of the research paper

development, it was apparent that Predictive Analytics is still an emerging term. While a simple

internet search on the terms “Big Data” and “Data Mining” generated many hundreds of peer-

reviewed academic papers from reputable online libraries such as IEEE Xplore and ACM DL, a

query on “Predictive Analytics” yielded only marginal amount of search result in terms of

volume and subject variety.

With this in mind, the aim of this paper is to solidify our understanding by surveying the

current landscape of Predictive Analytics in Data Mining with Big Data. The research result

would contribute to the knowledgebase of the evolving field of Predictive Analytics. In doing so,

this essay attempts to support the survey with contemporary best practices, empirical

experiments and case studies to shed light on this embryonic discipline. Also, the research result

would indicate theories, methodologies, models and tools that are commonly adopted by

researchers and frequently used by businesses. Thus, the research covers the spectrums in both

general application and specialized academic research domains.

Page 14: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 14

14

Research Problem

Background

The need to predict future or to explain pattern of natural phenomena bear many

implications to human understanding of epistemology. To anticipate an event, is by preparing a

response to a possible outcome, which is to say, judging from what have been known and infer

them to the unknown. The advantages of the ability to look ahead to yet-to-occur events are

plentiful, the following are a few examples of how the application of prediction allows us to

avoid unfavorable impending outcome:

Improves customer satisfaction with personalized purchasing recommendations;

Improves business competitiveness by being able to foresee customer behavioral

changes;

Avoids systematic economic downfall by assessing consumer credit risk scores;

Predicts earthquake by detecting and analyzing seismic activities;

Supports agricultural planning by forecasting weather;

Predicts seasonal influenza virus strain.

For consumers, a recommender system based on association rules analysis, advises on

merchandise purchase suggestions that might be of interest to online customers. The goal is to

improve overall customer satisfactions in reducing the time to manually locate interesting items.

This might seems trivial, however, when Predictive Analytics improves the online purchasing

experience of a consumer, the ripple effect could encourage more people to make more online

purchases rather than purchasing from brick and mortar stores. Thereby, the online shopping

environment eliminates the need for consumers to travel with transportation. Collectively, this

could lead to the reduction in traffic accidents and decreases CO2 emission footprint due to

Page 15: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 15

15

lower overall usage of automobiles. Therefore, Predictive Analytics can exert impactful

cascading changes on individuals that can lead to societal changes. The positive environmental

impact as a result of human behavioral change is only one of the many examples that Predictive

Analytics can deliver.

Predictive Analytics, strictly speaking, is a subset of Data Mining field which is part of

the Data Science discipline. The term Analytics itself derives from the science of data analysis

that is commonly associated with another term Business Intelligence to describe the provisioning

of decision support in businesses. Predictive Analytics is supported by applying mathematical

and statistical techniques to derive meaning from data and systematically find patterns in data for

decision making directives. The applications of Predictive Analytics range across both the

academia and the industries. The relationship between Business Intelligence and Data Ming is

depicted in Figure 1.

Page 16: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 16

16

Business

Intelligence

Data MiningData

Warehousing

Data

Aggregation

Data

Cleansing

Analytics

Processing

Data

Normalization

Figure 1: Business Intelligence Taxonomy1

To use Predictive Analytics, is to apply mathematics, statistics and probability theory in

conjunction with the overarching computer science discipline of machine learning, data

modeling and algorithm development.

The Problems

Predictive Analytics has a broad scope and wide application. Predictive Analytics cannot

be performed in isolation and should be considered a systematic approach for systems to work in

unison to derive result. To that end, Predictive Analytics at its core is realized by machine

learning and substantiated by statistical techniques to model, analyze and deduce knowledge.

The extracted data is expected to contain actionable information from seemingly uninteresting

dataset. Predictive Analytics mandates the use of database systems to maintain a collection of

1 The diagram depicts the relationships between various components of Data Mining and Data

Warehousing in relation to Business Intelligence. Each arrow represents a specific hierarchical relationship between

a set of given two entities. The entity receiving the head of an arrow is the sub-entity of its parent entity which

contains the tail of an arrow.

Page 17: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 17

17

data for processing and uses analytical modeling to transform raw data into actionable

knowledge.

To support the initiative of applying predictive analytics, a number of issues in relation to

data management and computational resources must be addressed. From an infrastructure point

of view, what network components2 and standards3 are required to support Predictive Analytics?

From an architecture point of view, what software and services4 are needed to construct and

support Predictive Analytics? From a conceptual standpoint, what are some of the principles,

issues and risks5 involved in Predictive Analytics, especially with regards to ethical concerns?

Within the context of Data Mining, what are some of the best practices6 in integrating Predictive

Analytics in both academia and business settings? Finally, how does Big Data fit into the

equation and how Big Data augments the infrastructural, architectural and conceptual paradigm

in our collective understanding of information management?

While Cloud Computing is part and partial to Big Data application, it will not be the

focus of this paper and will only be briefly discussed within the domain of Predictive Analytics.

The Definitions of Predictive Analytics and Big Data

Gartner Research’s definition of Big Data is widely adopted; the three Vs of Big Data

consists of Volume, Variety and Velocity (Gartner.com, 2014). While the three Vs definition of

Big Data is prevalent and distinguishes the unique aspects of Big Data, it does not serve a

suitable description for Big Data within the context of Predictive Analytics. As such, this paper

2 Refer to the Cloud Computing section for more information. 3 Refer to The Predictive Model Markup Language (PMML section for more information. 4 Refer to Predictive Analytics Performance Optimization section for more information. 5 Refer to CHAPTER V for more information. 6 Refer to CHAPTER III.

Page 18: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 18

18

attempts to provide a definition of Big Data that adequately describe its properties within

Predictive Analytics.

Big Data: The available and accessible set of historical data and metadata that capture

seeming unrelated multidimensional facts and events based on the result of human-created and

machine-generated actions.

Given the above definition of Big Data, the definition of Predictive Analytics is defined

below.

Predictive Analytics: To maximize the signal-to-noise ratio through the analysis of Big

Data. To use the result of such analysis in combination of the advanced techniques of statistical

modeling and the assistance of high performance computing devices, to derive meaningful

information that provide a higher-than-guessing accuracy and precision. The derived

information is capable of predicting trends and the validated result of each prediction will be

used in updating the underlying statistical model continuously and perpetually.

Organization of the Remaining Chapters

The approach in this paper will take the path of conducting literature review in the

context of Predictive Analytics, Data Mining and Big Data independently and collectively as a

single problem domain. Meta-analysis7 will be employed throughout this paper to contrast

conceptual constructs proposed by different researchers and to identify overlapping areas of

concerns.

Starting with CHAPTER II, this essay attempts to provide an in breadth coverage of the

current landscape of innovation in research and development with regards to the subject matter.

7 Meta-analysis is a type of research method for producing new information by contrasting the results from

previously related researches. Conducting meta-analysis does not require designing and performing an actual

experiment.

Page 19: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 19

19

CHAPTER II contains mainly literature reviews of existing research papers that discussed the

subject matter of Predictive Analytics in Data Mining with Big Data. The chapter would provide

readers with a rudimentary understanding of the subject matter. In the later part of CHAPTER II,

it progresses onto more specifically focus research areas such as intra DBMS modeling8.

CHAPTER III highlights the practical use of Predictive Analytics across many domains.

CHAPTER IV deals with methodologies, methods, best practices that are employed by

researchers and practitioners within the domain of the subject matter. CHAPTER V discusses

issues, challenges and recent trends in both academic and commercial spaces. Finally,

CHAPTER VI summarizes the conclusions and recommendations resulted from this study. Some

appendixes are also attached at the end of this report to provide additional material that supports

our analysis of the different aspects presented in different chapters.

8 Refer to the Predictive Analytics Performance Optimization section for more information.

Page 20: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 20

20

CHAPTER II

REVIEW OF RELATED LITERATURE

Introduction

The research topic in question is an area of intense interest from researchers and

practitioners. The expectation of Predictive Analytics to deliver value is a promising proposition.

Being able to forecast trends and to predict future behaviors has many implications ranging from

disease control to credit risk scoring as they relate to a particular population or individual. To

understand Predictive Analytics, one must first understand the concepts and usage of Data

Mining as it pertains to the seven steps of Knowledge Discovery From Data (KDD). They are,

data cleaning, data integration, data selection, data transformation, data mining, pattern

evaluation and knowledge presentation (Han, Kamber, & Pei, Data Mining: Concepts and

Techniques, Third Edition, 2011). As such, the promise of Predictive Analytics hinges on Data

Mining to provide meaningful dataset in order to fully appreciate its predictive prowess.

As more people are starting to understand the benefits and values that Predictive

Analytics can bring, the trend to higher adoption of Predictive Analytics is becoming evident.

The statistics shown in (Waller & Fawcett, 2013) suggested an increasing popularity in the space

driven by the higher business competition level in recent years. The recent uptake in the interest

of Big Data also fused the interest in Predictive Analytics as the two domains are closely related.

Predictive Analytics and Data Mining are part of the data science field (Waller &

Fawcett, 2013); Predictive Analytics is considered a subset of Data Mining field due to the

logical order of KDD operation. Data Mining takes precedence in this respect as it identifies

existing patterns and trends in seemingly unrelated and uncorrelated data. Performing Data

Mining is done in a descriptive way.

Page 21: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 21

21

The paper by (Haas, Maglio, Selinger, & Tan, 2011) explained descriptive analytics in

relation to Data Mining as a last resort for decision-making. This is because descriptive analytics

only describes the present conditions, whereas predictive analysis is a model-driven and data-

driven approach for generating what-if scenarios which fully exploits the nuances of underlying

data. Given this description, the logical order of decision support process is show in Figure 2.

𝐷𝑒𝑠𝑐𝑟𝑖𝑝𝑡𝑖𝑣𝑒 → 𝐸𝑥𝑝𝑙𝑎𝑛𝑎𝑡𝑜𝑟𝑦 → 𝐸𝑥𝑝𝑙𝑜𝑟𝑎𝑡𝑜𝑟𝑦 → 𝑃𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑣𝑒 → 𝑃𝑟𝑒𝑠𝑐𝑟𝑖𝑝𝑡𝑖𝑣𝑒 (𝐷𝑒𝑐𝑖𝑠𝑖𝑜𝑛)

Figure 2: Decision Support Process

The relationship between the different models is shown in Figure 3. In this sense,

Descriptive Analytics takes precedence over Predictive Analytics. Predictive Analytics relies on

Descriptive Analytics to provide the descriptive information as well as a foundational framework

in order for it to function and be effective.

The paper by (Haas, Maglio, Selinger, & Tan, 2011) introduced the role of Prescriptive

Analytics in decision making. Prescriptive Analytics is supported by Predictive Analytics which

in turns supported by the deterministic and stochastic optimization techniques such as Decision

Tree method and Monte Carlo method, respectively. To that end, Predictive Analytics produces

what-if scenarios for Prescriptive Analytics to derive objectives focus and constraints balanced

decisions. Thus, Prescriptive Analytics is logically separated from Predictive Analytics and

Prescriptive Analytics is positioned as a post-process of Predictive Analytics for final decision

support. In other words, the order of operation is described as 𝑋 → 𝑌 → 𝑍 where 𝑋 is Descriptive

Analytics, 𝑌 is Predictive Analytics and 𝑍 is Prescriptive Analytics.

Page 22: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 22

22

Descriptive

Analytics

Predictive

Analytics

Explanatory

Modeling

Predictive

Modeling

Analytics

Processing

Exploratory

Modeling

Figure 3: Analytics Processing Taxonomy9

Relying on only Descriptive Analytics in Data Mining to extrapolate future outcome can

be disastrous as illustrated in the example provided in (Haas, Maglio, Selinger, & Tan, 2011).

The example of “Extrapolation of 1970-2006 median U.S. housing prices” chart demonstrated

the effect of a shallow prediction (i.e. extrapolation) versus deep predictive analytics where the

extrapolated housing prices beyond year 2006 was widely different than the actual prices

recorded between the year 2006 and 2010.

While Descriptive Analytics provides a simple picture of linear relationship between past

and future events, it is however not a solid predictor for the future. That is to say, Descriptive

Analytics can help in situations where extrapolation and interpolation are acceptable but

prediction using extrapolation is destined to fail as the margins of probable errors are not

factored in the equation. The main differences between Predictive Analytics and Descriptive

9 Each arrow represents the logical order between each component. For instance, Predictive Analytics

depends on the result from both the Exploratory Modeling and Predictive Modeling components.

Page 23: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 23

23

Analytics lie in the application of advanced techniques, some of which are, machine learning,

advanced modeling, causality approximation, root-cause analysis, sensitivity analysis, model

validation and many other probability driven model designs. As such, Haas et al. concluded that

the focus for Business Intelligence should operates more with deep Predictive Analytics rather

than shallow Descriptive Analytics. The standpoints of model and data need to have a balanced

perspective and must not overly emphasize solely on the part of either the model or the data

when performing analysis.

Explanatory versus Predictive Modeling

The concerns brought by (Haas, Maglio, Selinger, & Tan, 2011) are shared by Shmueli et

al. in (Shmueli & Koppius, 2010). The authors explained how the common approach of

empirical models for explanation (i.e. explanatory statistical model) is fundamentally distinct

from the empirical models for prediction (i.e. empirical predictive model). One of the key

distinctions between the aforementioned two models is the difference in analysis goal.

Explanatory model is used for testing causal hypotheses while predictive model is ideal for

predicting new observations and assessing predictability levels.

Explanatory model and predictive model can be complementary in certain regards;

however they differ from each other in terms of achieving the end goal at an empirical level,

because they are used for different purposes and in different contexts. The notion of causality

between theoretical constructs is deeply rooted in the explanatory model, whereas predictive

model depends on the associations between measureable variables (Shmueli & Koppius, 2010).

In the context of empirical modeling, the selection of statistical modeling methodology

and the expectation of application determine the techniques employed. Principally, the

dichotomy of explanatory modeling and predictive modeling leads to a contrasting outcome that

Page 24: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 24

24

affects model selection, model techniques, model evaluation and validation. The result of these

differences was summarized in Table 1.

The Role of Predictive Model

The role of Predictive Analytics in scientific research differs from Explanatory Analytics

as well. Predictive modeling shares the principles of grounded theory in research design, in

which, the main purpose is to generate theory from both qualitative and quantitative data. In fact,

predictive modeling is theory generative in nature rather than theory derivative. Like grounded

theory, predictive modeling attempts to produce a theory (i.e. a prediction) based on observations

(i.e. data).

Another role of the predictive model is to assess the predictive power of any predictive

model based on the measure of distance between theory and practice (Shmueli & Koppius,

2010). Explanatory models are usually constructed with in-sample dataset that serves as a

training set and are validated using out-of-sample dataset from the same sample set. Models

created under this methodology can be rigid and bias towards the sample dataset and unable to

account for rapid shift in data context due to a possible model overfitting, which diminishes their

predictive power. Therefore, the theory generative properties of predictive modeling are more

suitable to be used to infer to probable outcomes through measureable variables where causality

plays a lesser role to probability.

Underlying Models Differences

In fact, the result of applying explanatory techniques to predictive modeling would

inevitably defeat the original goal and diminishes the predictive power of a chosen model. The

key point in the two modeling approaches is the type of relationships being measured as well as

the external factors themselves. The author in (Shmueli, 2010) defined the terms as follow:

Page 25: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 25

25

Explanatory Modeling: A retrospective approach to explain the underlying

causal relationships between theoretical constructs that is both descriptive and

explanatory. This definition is depicted in Figure 4.

Predictive Modeling: A prospective approach to predict unknown outcome based

on the association relationships between measureable variables. This definition is

depicted in Figure 5.

Cause Construct Effect ContructCause-effect

Figure 4: Explanatory Model

Measurable

Variable X

Measurable

Variable YAssociation Relationship

Figure 5: Predictive Model

Figure 4 depicts the basic premise of an explanatory model, in that, the cause and effect

constructs are linked by a unidirectional cause-effect relationship. In statistical terms, the cause

construct can be expressed as an independent variable (i.e. factor) and the effect construct can be

realized as dependent variable (i.e. outcome). Empirically, it is important to point out that

Page 26: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 26

26

correlation does not equate to causation. Variables that are correlated, either positively or

negatively correlated, can be affected by externally latent variables rather than the observable

variables. For example, if variable 𝑋 and 𝑌 are correlated with a correlation coefficient (i.e.

𝑅(𝑋,𝑌)) value of 0.9, it can be both influenced by a third variable 𝑍 that directly manipulates

variable 𝑋 and 𝑌. However, 𝑍 was not detected because 𝑍 is not an observable variable. In other

words, had the effect of variable 𝑍 not been present, the correlation between variable 𝑋 and 𝑌

will no longer be observed and the perceived correlation between the two variables 𝑋 and 𝑌 is a

direct result of a third variable 𝑍. Therefore, only variable 𝑍 has a direct causal effect on variable

𝑋 and 𝑌. However, without the knowledge of 𝑍, the illusion was perceived as if variable 𝑋 and

𝑌 are causally related.

Explanatory Model allows researchers to measure the model validity of a given set of

hypotheses. Methods for measuring model internal validity can be achieved through R-squared

and p-value statistical calculations that validate and evaluate variables causality and statistical

significance. The result can be used to either accept or reject a particular hypothesis for a

quantitatively explanation on a given phenomenon. However, the power to explain within a

historical context for which explanatory model was designed to perform, does not extend to

predictive application which operates with little regard for causality, but it operates with a higher

focus on measurable outputs.

Predictive Model, on the other hand, is determined by a loose association between

measureable variables within an observable relationship that does not denote any direct causal

meaning. This relationship is shown in Figure 5. This key difference enables predictive model to

be highly adaptive as it meets the goal to achieve high predictive power rather than explanatory

power.

Page 27: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 27

27

Explanatory modeling methods such as simple regression-type methods are not suitable

for predictive tasks as described in (Shmueli & Koppius, 2010) and (Shmueli, 2010). A

collection of Descriptive Analytics techniques are presented in Figure 6 based on the survey

result. To improve predictive power, methods involving machine learning and probability

calculation often yield better results, some the techniques are shown in Figure 10. Predictive

model is not intended for past-conformity as there exists a risk of model overfitting for model

that is tightly coupled to historical events. For predictive model to be able to foretell an event, a

level of tolerance must be built into the model in order to adequately handle unforeseen events.

Descriptive

Analytics

Data Cube

Technology

Online

Analytical

Processing

Linear

Regression

Modeling

Frequent

Patterns

Apriori

Algorithm

Frequent

Pattern

Growth Tree

Least

Squares

Regression

R Square

Measure

Figure 6: Descriptive Analytics Taxonomy

Page 28: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 28

28

The following table summarized the arguments presented in (Shmueli & Koppius, 2010)

and in (Shmueli, 2010):

Table 1: Explanatory Model versus Predictive Model

Explanatory Model Predictive Model

Characteristic Descriptive and Explanatory Forward-looking

Targeted Relationship Causation Association

Targeted Relationship Theoretical Constructs Measurable Variables

Perspective Retrospective Prospective

Data Continuity Continuous changes in data Continuous and discontinuous changes in data

Data Bias High Low

Data Volume Low High

Data Cleansing Requirement

High Low

Data Noise Undesirable Expected

Data Partition Less Common Common

Exploratory Domain Limited Wide

Exploration Interactivity

Low High

Data Dimension Reduction

High Low

Popular Analysis Models

Simple regression-Type Models Algorithmic Modeling such as neural networks and k-nearest-neighbors

Model Transparency High Low

Power Assessment Techniques

R Square and F-Test Backtesting

Data Mining and Big Data

Data Mining is an all-encompassing term to describe the methodology, strategy and the

common approaches to extract knowledge out of raw data using techniques from many domains

shown in Figure 7. Performing Data Mining is to put data into context, to produce information

that is relevant and useful for solving specific problems. Very often, these problems are concerns

that crosscut many academic and business areas, making the multidisciplinary field of Data

Mining an indispensable one for solving problems using an evidence (i.e. data) based approach.

With the improved understanding of data through Data Mining, the need to predict future

Page 29: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 29

29

occurrences based on historical references prompted the development of Predictive Analytics.

Predictive Analytics forecasts trends and behavior patterns augmented by the three properties of

Big Data (i.e. volume, variety and velocity).

Figure 7: Data mining adopts techniques from many domains (Han, Kamber, & Pei, Data Mining: Concepts and Techniques,

Third Edition, 2011)

The Big Data Problems

The challenge in the understanding of Big Data stems from the nebulas term Big Data

coined by many researchers to describe the accelerating data volume since the dawn of the internet

age.

The publicly accessible internet data have been increasing in size at an exponential rate as

far back as 1995 as shown in (Leiner, et al., n.d.) and in (Lesk, n.d.). Between the year 1999 and

2001, the estimated number of internet web pages has grown from less than 1 billion to more than

4 billion pages in a period of 2 years (Murray & Moore, 2000). This rate continued to accelerate,

the estimated size of the index-able internet as of 2011 is approximately 5 million terabytes or 5

Page 30: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 30

30

exabyte of data (Clair, 2011). The ever increasing data volume managed by organizations is putting

pressure on storage capacity. In order to meet the demand, a different strategy is in order to deal

with many of these challenges.

The European Bioinformatics Institute (EBI), part of the European Molecular Biology

Laboratory (EMBL), had gathered approximately 28 petabytes worth of bioinformatics data as of

this writing (About Us Background, 2014) to assist tasks such as DNA sequencing, drug resistance

testing and drug R&D. The large dataset spans many different types, spanning genes, proteins

expressions, small molecules and protein structures.

The European Organization for Nuclear Research (CERN) faces the same storage capacity

and data volume challenge with an annual growth of 15 petabytes worth of data for physicists to

sift through (Computing, 2014). The data are generated by the Large Hadron Collider (LHC) at a

rate of 600 million times per second in an effort to recreate the moment immediately following the

big bang event based on the Big Bang Theory. Each particles collision set forth a chain reaction

that led to a series of complex events. The exponential growth of each collision and the subsequent

reactions created massive amount of data to be persisted in CERN Data Centre (DC) for physicists

to analyze.

Furthermore, the Low Frequency Array (LOFAR); a radio interferometer for detecting low

radio frequencies between the range of 10 MHz and 240MHz, collects data in petabytes scale per

year with each single file approaching terabyte sizes (Begemana, et al., 2010) for geoscience and

agricultural applications.

The sheer volume of internet data also gave raise to data complexity and variety in terms

of data structure, data sources and data types. Researchers have devised various strategies to

collect and process such tremendous amount of data in both data volume and data variety

Page 31: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 31

31

aspects. Some of the strategies involve storing unstructured data in distributed NoSQL databases

and use MapReduce method for job-based distributed data computation. In this domain, the

Apache Hadoop Platform (Welcome to Apache™ Hadoop®, 2014) is currently dominating the

research and commercial space.

The NoSQL Solution

The statistical algorithms that underpin many of these strategies will be discussed in

CHAPTER IV. Many of these strategies built on the traditional disciplines of mathematics and

statistics, created an overall interdisciplinary approach to tackle the challenges brought by Big

Data. As such, the computer science discipline plays an important role as researchers continue to

innovate using an interdisciplinary approach that combines mathematics, statistics and software

engineering. The advancements in database technologies and distributed computing are some of

the great examples that combat the three Vs problem of Big Data based on the centuries-old

principle of divide-and-conquer.

Applying Data Mining in Big Data is a complex process that involves many procedures

and is poised to uproot our collective understandings of the most fundamental component in

information management, the relational database technology. As such, this gave rise of the

NoSQL database technology (Russom, 2011) (Menegaz, 2012) (Chen, Chiang, & Storey, 2012)

that was designed to better handle the problem of the three Vs of Big Data. NoSQL is a non-

relational data model approach designed to face the challenges presented by Big Data, that are,

volume, variety and velocity.

In short, NoSQL is designed with Big Data in mind that have made trade-offs to satisfy

the requirements on performance, size, transactional support and features. One such design

criteria led to the handling of semi-structured and unstructured data in a distributed environment

Page 32: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 32

32

for data management. The various implementations of NoSQL on the market today include

CouchDB, MongoDB, FlockDB and Apache Casandra, to name a few. Many of these

implementations rely on the Hadoop (The Apache Software Foundation, 2013) software library

which includes common NoSQL utilities, distributed file system, job scheduler and cluster

management as well as the centerpiece component MapReduce.

The MapReduce method simply describes the procedure of work distribution (i.e. map) to

computing nodes and result aggregation (i.e. reduce) from the nodes. The MapReduce method

was designed for distributed and parallel data processing, an important strategy to handle the

overwhelming volume of data and rapid data creation that Big Data imposes.

Note that there is no single or unified way to implement NoSQL solution. However, many

existing implementations can be categorized into the following three NoSQL database types: Key-

Value Store, Graph Database and Document Store (Han, Kamber, & Pei, 2011). Each

implementation of these database types had made trade-offs to balance the various concerns. These

concerns are characterized by their respective advantages and disadvantages. For instance, Graph

Database is optimized for acyclic and cyclic graph objects while Key-Value Store represents data

in an unstructured key-value pairs for flexible data representation.

The common theme is that they are based on non-relational data model and distributed

computing focus. Also, most of the NoSQL solutions today are unstructured data focus and they

are not full ACID (i.e. Atomicity, Consistency, Isolation and Durability) compliant (Han, Kamber,

& Pei, 2011) which is the one of the most important features in today dominated relational data

model.

Page 33: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 33

33

The Apache Hadoop Platform

The Apache Hadoop Platform composes of the two key components: a distributed file

system called HDFS for distributed data management and data storage, as well as, the

MapReduce method for distributed data querying and data computation. There are other

supplementary technologies in the Apache Hadoop Platform including Apache Sqoop. However,

they are not the focus of our discussion in relation to our research topic of Big Data.

HDFS stands for Hadoop Distributed File System and it was designed to run on

commodity hardware. This design goal allowed the operation of HDFS to be highly scalable and

available. To the end, HDFS provides built-in support for data fault-tolerance (i.e. data

replication) and load-balancing (i.e. MapReduce). In other words, HDFS is a single logical file

system distributed across many data servers and it is able to scale on demand based on required

capacity. A high-level HDFS architecture is shown in Figure 8.

Figure 8: HDFS Architecture (HDFS Architecture Guide, 2014)

Page 34: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 34

34

Given the architecture shown in Figure 8, MapReduce provides a means to query data

across the disparate data servers. MapReduce, in a nutshell, consists of the following two

functions:

The Map function is performed by the master node to partition and distribute

function input into multiple small inputs to be handled by downstream worker

nodes. That is, 𝑓(𝑥, 𝑦) → [

𝑥1 𝑦1

𝑥2 𝑦2……

] where (𝑥, 𝑦) is the key-value pair input of the

Map function and the matrix is the output of the breakdown of original input.

The Reduce function is also performed by the master node to aggregate and

collect outputs from the disparate worker nodes. That is, 𝑓([𝑥1

𝑥2

…] , [

𝑦1

𝑦2

…]) → 𝑧 where

𝑧 is the result of the aggregation of the individual outputs from the worker nodes.

Both HDFS and MapReduce are low level components of Hadoop. Of course, performing

operations directly against low level APIs are time-consuming and unproductive. Without the

introduction of some higher level software layer, it would be difficult to drive Hadoop adoption.

For this reason, the burgeoning NoSQL based DMBS that are designed to operate on top of the

HDFS becomes an indispensable component of any Hadoop implementation.

Recent Hadoop development includes subprojects such as Apache Accumulo, Apache

Bigtop, Apache Chukwa, Apache Sqoop and Apache Flume to improve the overall capability of

Hadoop platform. Most notably, the Hadoop YARN project is a MapReduce alternative to data

processing.

Page 35: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 35

35

NoSQL Data Model versus Relational Data Model

The object structure in relational data model is rigid. For instance, a relation (i.e. table) is

defined with a set number of attributes (i.e. columns) as well as a collection of tuples (i.e. rows)

representing each instance of an object with matching attributes. Relationship, in this case, is

represented by an additional link attribute (e.g. foreign key) introduced in a derived relation (e.g.

child table). A rigid model such as relational model has many merits, one of which is a higher

level of entity and referential integrity insurance. Since each tuple in a relation is guaranteed with

a set number of attributes, constraints (e.g. not nullable and foreign key constraint) can be set in

accordance to the level of data integrity required to ensure that data at each tuple level will

adhere to the defined constraints.

The NoSQL model rivals the traditional relational data model in that trade-offs have been

established in favor of higher performance and scalability. Some of the NoSQL implementation

includes, but not limited to: Apache Cassandra, Apache CouchDB, Berkeley DB, FlockDB,

MongoDB, Neo4j and Object DB. All of the abovementioned NoSQL implementations fall under

one of the following three model categories:

Key-Value Stores: The content of the data is represented as a collection of

individual key and value pairs. Dynamo (DeCandia, et al., 2007) is one of the

many implementations under this NoSQL model.

Document Store: The content of the data is organized in a per document

container object. A single document (e.g. XML and JSON) encapsulates all

attribute data for a given object. Couchbase Server (Couchbase.com, 2012) is one

of the many implementations under this NoSQL model.

Page 36: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 36

36

Graph Database: The content of the data are represented in graph objects based

on Graph Theory. The object attributes are described in nodes and edges that are

interconnected with other object attributes. AllegroGraph (Shimpi & Chaudhari,

2012) is one of the many implementations under this NoSQL model.

Contrary to relational data model, the above NoSQL model structures are not rigid. As

such, they are often referred as unstructured data models. The flexibility exhibited in NoSQL

data model is highly desirable in Big Data application, due to the fact that, Big Data is inherently

unstructured and diverse. NoSQL data model provides a foundational framework for managing

Big Data where Predictive Analytics can efficiently perform actions against disparate and

distributed databases.

Predictive Analytics versus Other Forecasting and Mining Methods

The fields of Big Data and Data Mining have been evolving to bring new ideas and

innovations to the field of information management. Though, the term and application of

Predictive Analytics in Data Mining came before the term Big Data. Performing analytics,

predictive or not, predated both Big Data and Data Mining. This is partly due to the fact that the

underpinnings of analytics stem from some of the traditional disciplines, some of which are,

mathematics and statistics (Han, Kamber, & Pei, 2011).

The application of analytics has far reaching impact to business bottom-line. Businesses

often need to find patterns in their customer behavioral data in order to derive strategies to

improve business prospect as they better their understanding of the customers. A

multidisciplinary and interdisciplinary approach to uncover business intelligence began to

emerge. This translates into tangible practices of risk management, outlier detection, web

analytics, logistics management and business optimization. As Big Data gains popularity, the

Page 37: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 37

37

underlying promise of Predictive Analytics remains the same. However, given the vast amount of

data and wide variety of data available at disposal, combined with the fact that data are arriving

at near real-time speed, approaching knowledge extraction with novel methods is in order.

Human has long devised numerous techniques to predict and anticipate different

outcomes. Industries such as the insurance industry are highly dependent on advanced predictive

techniques for their business. This gave rise to an entire discipline dedicated to the study of risk

and uncertainty, the actuarial science. Insurance companies employ actuaries to calculate

insurance premium based on various factors as to forecast population and assign risk scores to

groups of individuals. This deterministic model based on the predetermined segregation of

population sample is achieved through the same underpinning techniques as Predictive Analytics.

However they vastly differ in methodology and approach, which highlighted one of the major

differences between actuarial forecasting and Predictive Analytics.

Actuarial forecasting approaches problems from a top-down fashion and addresses

questions that can be answered for a predetermined chance and probability. Predictive Analytics,

on the other hand, answers questions through a bottom-up approach at individual level (Siegel,

2013). For example, retail business would be interested in calculating the likelihood of customer

𝐴 to purchase item 𝑌 when item 𝑋 was purchased during 𝑇 hours. Determining 𝑌 in {𝑋, 𝑌} for

every 𝐴 and 𝑇 leads to actionable outcome for which basket analysis is done via association

rules (Han, Kamber, & Pei, 2011). Predictive Analytics provides actionable decision-support

information that can benefit targeted marketing by looking at data gathered at individual level – a

bottom-up theory generative approach. It is in fact the granularity of data itself that became the

enabler for Predictive Analytics.

Page 38: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 38

38

Conceptually, Data Mining is a superset of Predictive Analytics as shown in Figure 9.

From a methodology perspective, Predictive Analytics is a term describing the principles and

techniques used in Data Mining specifically for predictive analysis. The Data Mining

methodologies includes characterization and discrimination, frequent patterns, associations,

correlations mining, classification analysis, regression analysis, cluster analysis and outlier

analysis, as shown in Figure 9. A detail taxonomy for Predictive Analytics is shown in Figure 10.

Classification and

Regression

Analysis

Mining Frequent

Patterns,

Associations, and

Correlations

Characterization

and

Discrimination

Cluster Analysis

Outlier Analysis

Predictive

Analytics

Figure 9: Predictive Analytics in Data Mining

Page 39: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 39

39

Predictive

Analytics

Nonlinear

Regression

Decision

Tree

Artificial

Neural

Network

Classification

Machine

Learning

Bayesian

Belief

Networks

Naïve Bayesian

Calssification

Cluster

Analysis

Logistic

Regression

K-nearest

Neighbours

Supervised

Learning

Unsupervised

Learning

DBSCAN

Stochastic

Gradient

Descent

Isotonic

Regression

Rule Based

Classification

Backpropagation

Artificial Neural

NetworkSupport

Vector

Machine

Figure 10: Predictive Analytics Taxonomy

The techniques in Data Mining shown in Figure 9 are applicable to Predictive Analytics

for predictive analysis shown in Figure 10. However, Predictive Analytics has a strong

preference in employing the classification and regression analysis methodologies which will be

Page 40: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 40

40

discussed in CHAPTER IV. Many popular implementations such as decision tree and artificial

neural network fall under classification and regression analysis category. The rule-based

classification methodology is particularly suited for certain Predictive Analytics tasks since the

class/label identification and prediction-to-class/label mapping processes are initiative and easily

explicable. The regression methodology is common in numerical based analysis to produce a

mathematical function that best describes the data. Other techniques in Data Mining are also

used during predictive analysis, however, they typically points to usages involving data

preprocessing that deals with descriptive and explanatory type analysis.

The Predictive Model Markup Language (PMML)

PMML is a standard markup language for Data Analytics. The plethora of techniques

used in Data Mining and Machine Learning exert pressures on model interoperability between

applications and services. The XML based Predictive Model Markup Language (PMML) was

created by the independent and vendor-led consortium called the Data Mining Group (DMG).

The DMG standardizes model file format containing model definition (Data Mining Group,

2014). Since the inception of PMML version 0.7 in July 1997, the DMG group has been

continuously improving the model definition coverage defined by the PMML schema. The most

up-to-date version of PMML is version 4.1 published on December 2011 with enhanced post-

processing capabilities and new model elements amongst other updates. An example of the

generated PMML code can be found under the PMML Code Example section of APPENDIX A.

PMML Adoption

As of this writing, there are over 30 members from data analytics industry and

government organizations have adopted the PMML model. Also, many modeling applications

have adopted PMML. Some notable modeling applications are: IBM SPSS (IBM, 2014),

Page 41: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 41

41

KNIME (KNIME, 2014), R (R, 2014), RapidMiner - Extension (RapidMiner, 2014) and Weka

(Weka, 2014). Many dataset repositories are also adopting PMML and are offering sample

datasets in PMML format. In fact, the UCI Machine Learning Repository currently hosts

approximately 300 datasets (UCI Machine Learning Repository, 2014) for model development

and testing, many of which are in PMML standard format.

PMML is gaining adoption in both academia and business domains. The reflection in

PMML adoption suggests a convergence of modeling techniques and the stabilization of

competing predictive modeling methodologies for Predictive Analytics. In fact, the Model

section of the PMML schema implements some of the standard modeling methods such as

Support Vector Machine (i.e. SupportVectorMachineModel) and Decision Tree (i.e.TreeModel).

In order to define a XML schema that is interoperable, the elements in XSD that describe the

underlying models must contain the model structure and the composite parts. The PMML schema

thus echoes a general consensus amongst applications that support the common predictive

models, further hinting to the maturity of the Predictive Analytics field.

As noted in (Guazzelli, Jena, Lin, & Zeller, 2011), the main goal of DMG is the

development of PMML, and now aims to make PMML a de facto standard to represent Data

Mining models. Many researchers share the same view of embracing PMML and understand its

benefits to the research community (Guazzelli, Stathatos, & Zeller, 2009). As a standard, PMML

benefits also the Cloud Computing community through promoting interoperability and openness.

PMML Document Structure

Under the PMML specification, the components that made up the model are defined

under the following major sections (DMG, 2014) as shown in Figure 11. The individual parts of

a PMML document are defined to specifically describe a subset of an overall model. The PMML

Page 42: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 42

42

document structure ensures the interoperability between supporting modeling applications by

means of a well-defined schema and use of XML human readable open data format.

Figure 11: PMML Schema (DMG, 2014)

Header Section: The header section of the PMML schema contains model

metadata of copyright information, model description, generator application,

name, version, annotation and timestamp. An example of a simple header section

can be found at the PMML Code Example - Header Section.

DataDictionary Section: The DataDictionary section contains the number of data

fields and field definitions such as data type, data range, available data values and

data field category. An example of DataDictionary section can be found at the

PMML Code Example - DataDictionary Section.

TransformationDictionary Section: The TransformationDictionary section

contains general model preprocessing conditions that describe how data will be

Page 43: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 43

43

transformed from the original state into a desired state. The types of data

transformation PMML supports are normalization, discretization, value mapping,

functions and aggregation. An example of TransformationDictionary section for

value mapping transformation can be found at PMML Code Example -

TransformationDictionary Section.

Model Section: The Model section is represented by MODEL-ELEMENT group

under the PMML 4.1 schema with the main purpose of describing various

modeling techniques used in research and analysis. This is the core section of the

PMML 4.1 standard as most intra-model information is contained within the

model section. The information included within this section varies greatly due to

the differing terminologies and underlying concepts across modeling techniques.

The latest version supports the following models: AssociationModel,

BaselineModel, ClusteringModel, GeneralRegressionModel, MiningModel,

NaiveBayesModel, NearestNeighborModel, NeuralNetwork, RegressionModel,

RuleSetModel, SequenceModel, Scorecard, SupportVectorMachineModel,

TextModel, TimeSeriesModel and TreeModel. The element name unambiguously

represents the names of the modeling techniques. For instance,

SupportVectorMachineModel element is meant to capture model information for

Support Vector Machine (SVM) model. An example of Model section for SVM

can be found at PMML Code Example - Model Section – Support Vector

Machine.

Page 44: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 44

44

PMML Interoperability and Application

The Service Oriented Architecture (SOA) provides the foundational framework for web

services to interoperate based on common HTTP communication standards such as Simple

Object Access Protocol (SOAP) and Representational State Transfer (REST). Thus, the standards

achieve interoperability amongst participating services. The same holds true for predictive

modeling in a cloud computing environment where PMML is one of the key enablers to maintain

interoperability amongst modeling applications.

The PMML standard brings tangible benefits in modeling application interoperability,

improve collaboration amongst researchers and streamline workflow that involves the multistep

process of predictive modeling. In this regard, cloud computing such as Software as a Service

(SaaS) model (i.e. Google Apps, Amazon EC2) benefits greatly from predictive modeling with

PMML, for reasons that software interoperability and integration are critical in cloud computing.

The Amazon EC2 (i.e. Amazon Elastic Compute Cloud) enabled ADAPA scoring engine

is one such example that fully took advantages of PMML and SaaS model. The discussion in

(Guazzelli, Stathatos, & Zeller, 2009) used ADAPA as the key example. In the study, ADAPA

was deployed as a model verification system that operates within a cloud environment. The

predictive model was built on the PMML standard for describing model definition that aids the

modeling process from model design to verification as shown in Figure 12. The predictive

algorithms were expressed in PMML to be scored by multi-instances ADAPA engine hosted on

Amazon EC2 infrastructure.

Page 45: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 45

45

Figure 12: Life Cycle of Data Mining Project using ADAPA (Guazzelli, Stathatos, & Zeller, 2009)

The many benefits of utilizing PMML in conjunction with the ADAPA engine within the

Amazon EC2 environment had been illustrated by the El Niño Neural Network modeling

example in (Guazzelli, Stathatos, & Zeller, 2009). The inherit benefits (e.g. low startup cost,

distributed computing power, streamline management and robust APIs, etc.) of cloud computing

combined with the interoperability of PMML, streamlines the process of predicting modeling to

a much greater degree. This applies to every step within the process, from model verification to

modeling testing.

Of course, the application of PMML in a cloud platform does not limit only to model

verification as in the previously discussed ADAPA study. Note that PMML was designed to be

multi-purpose in the predictive modeling domain and many researchers have performed model

execution based on model described in PMML to derive predictions. So to consider PMML

merely as a model data persistence protocol is an understatement to its potential.

In (Das, Fratkin, Gorajek, Stathatos, & Gajjar, 2011), Das et al reaffirmed the view that

the portability and interoperability of PMML are bridging the gaps between all participants

whom involved in the data mining process. PMML is a conduit that links between cross-teams

Page 46: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 46

46

and stimulates inter-organization communication. Consequently, PMML reaches far beyond

other direct means of team collaboration. PMML fosters team communication and collaboration

that were lacking in the predictive modeling practice domain.

PMML Enabled Architecture

A clear example of how PMML bolsters team communication and collaboration is the

experiment conducted in (Das, Fratkin, Gorajek, Stathatos, & Gajjar, 2011). The experiment

involves an EMC Greenplum database which is a derivative of PostgreSQL database, to act as

the backbone for the in-database processing experiment. EMC Greenplum was selected in the

experiment for its Massively Parallel Processing (MPP) share-nothing database architecture that

supports SQL and MapReduce parallel processing.

Under the MPP architecture, a typical configuration of Greenplum database is a

collection of servers divided into two roles: the master host and the segment host. The master

host is often setup with redundant servers as its role is critical in that the master host is

responsible for listening to client queries, optimally allocating the queries to segment hosts based

on parallel query plan and returning processed results to the client application. The segment

hosts are responsible for the actual performance of the query allotted by the master host. The

above process marks the archetypal MapReduce operation. This architecture is depicted in Figure

13.

Page 47: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 47

47

Figure 13: EMC Greenplum MPP database architecture (Das, Fratkin, Gorajek, Stathatos, & Gajjar, 2011)

The MPP database architecture is modular and distributed in nature as it was designed to

support cloud computing and Big Data processing. The MPP architecture demands an open and

interoperable model information exchange format to fully exploit its parallel processing

potential. This led to the experiment in (Das, Fratkin, Gorajek, Stathatos, & Gajjar, 2011) where

the researchers deployed the El Niño regression model to Greenplum database for predictive data

processing tasks. The El Niño regression model was created in R and the model was

encapsulated and deployed in PMML format.

One of the key aspects of how PMML aided the deployment process involved creating

SQL functions as the query execution language for Greenplum and the dynamic mapping of SQL

function to a section of PMML definition. The researchers highlighted that the information

Page 48: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 48

48

contained within the PMML DataDictionary and MiningSchema sections can be effortlessly

mapped to the SQL function specification. The active mining fields in PMML can be mapped to

the SQL function input parameters while the predicted mining field in PMML can be mapped to

the SQL function output parameter. This mapping process involving PMML-to-SQL-functions

conversion can be automated as the PMML schema and the SQL function specification are

clearly defined to support the model execution. In (Das, Fratkin, Gorajek, Stathatos, & Gajjar,

2011), the SQL functions took the form of User Defined Functions (UDF) and they were created

to support the massively in-database parallel processing design by facilitating high performance

data selection and data querying.

Once the UDFs are created, the results from the execution of those UDFs are then

returned to the client application for the performance of the actual predictive analysis. This

architectural design maintained a high performance benchmark score. The authors in this

experiment not only highlighted the apparent advantage of using PMML in a natively PMML

compliant applications such as R, but also illuminated how the open format of XML and PMML

helped in binding and automating seemingly incompatible systems such as Greenplum and R.

The demonstration by the authors encourages more researchers and practitioners to adopt the

PMML standards across the entire predictive modeling lifecycle.

Cloud Computing

Cloud Computing is a term denoted to a collection of hardware and software services

supported by organizations to provide on-demand access to these resources on the internet.

According to (Vouk, 2008), Cloud Computing build on the success of preceding technologies

including Virtualization, Distributed Computing, Grid Computing, Utility Computing, Web and

Software Services.

Page 49: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 49

49

Many businesses have taken notice of Cloud Computing platform in recent years as a

way to take advantage of utility computing paradigm exemplified by the pay-as-you go payment

model and virtualized computing infrastructure. Cloud Computing primarily exists in one of the

following service levels: Infrastructure as a Service (IaaS), Platform as a Service (PaaS) and

Software as a Service (SaaS). Each progression from one level to the next built on the previous

level and cumulatively adds layers of computing services. Given this layered separation of

concerns, businesses can purchase computing services on demand rather than as a lump sum

investment, such as paying for a package software solution that often adds costs to integration

and service management without utilizing the full solution capability.

The Cloud Computing paradigm brings many advantages in the context of statistical

modeling with Big Data. Model execution often demands high power data processing servers and

high capacity databases to sustain the strenuous operations imposed by Big Data mining

algorithms. These requirements would also levy a substantial upfront investment in hardware and

infrastructure. Cloud Computing makes a good economical option for predictive modeling in this

case.

In (Kridel & Dolk, 2013), the authors took Cloud Computing to the next logical level in

SOA: The Predictive Analytics as a Service (PAaaS). The authors recognized PAaaS and Models

as a Service (MaaS) are a subset of Software as a Service (SaaS) model under the Service

Oriented Architecture (SOA). PAaaS offers Predictive Analytics modeling services that are

tailored to common modeling needs by considering two layers of analytics service, they are

called: deep structure layer and surface structure layer. Depending on the resources availability

within an organization, one of the two aforementioned layers in performing predictive analytics

Page 50: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 50

50

can be externalized to the Cloud Computing platform while retaining other analytics functions

within the organization.

In addition, the ADAPA engine described under the PMML Interoperability and

Application section of this essay illustrated a Predictive Analytics and Cloud Computing use

case. The interaction between the user and the cloud-enabled ADAPA system is supported by

both the web console mode and the web service mode. The former allows the end user to

remotely operate a GUI console through a HTTPS connection while the latter allows web service

API communication between programs and ADAPA. Using web service API also allows batch

processing to automate model verification process - from loading sample dataset to reporting

model score result. The cloud-enabled ADAPA system addressed the important concern of

security as well. The connections between all endpoints are secured via HTTPS and the dynamic

instancing nature of ADAPA means no residual information retained once the instances have

been terminated.

Dichotomy of Predictive Analytics Skillsets

The motivation behind the separation of the two layers of concerns stems from the deeply

divided human resource skillsets (e.g. Information management versus statistical modeling) and

their availability within an organizations. For many businesses, a successful implementation of

Predictive Analytics project is a great challenge. This is due to the dichotomy of key skills

between IT professional and modeling analyst.

The domain specific skillsets for IT professional and modeling analyst are distinct. The

different skillsets calls for very different qualifications. IT professionals are often responsible for

ensuring data and system availability (i.e. surface structure layer) while modeling analysts focus

on statistical modeling and predictive algorithm development (i.e. deep structure layer). This

Page 51: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 51

51

dichotomy creates a challenging proposition for small businesses to gain entry into the Predictive

Analytics space without the help of PAaaS to set the stage for two key cornerstones of predictive

analytics: infrastructure platform and modeling framework.

The chance of acquiring individuals with the proper skillsets becomes increasingly rare as

the available human resources decreases within an organization. Therefore, talent acquisition is

challenging for small organizations with respect to initiating Predictive Analytics project

compared to large organizations. This is why it is important to understand the logical separation

between skillsets that are required to deal with deep structure layer functions versus the skillsets

that are needed to maintain surface structure layer functions. To clarify, the deep structure layer

refers to a stable and foundational schema that rarely changes (Kridel & Dolk, 2013). The

surface structure layer, on the other hand, is considered as various instantiations of the deep

structure layer that exists in malleable and volatile forms (Kridel & Dolk, 2013).

To put this in a concrete context, a deep structure layer encapsulates model schema and

model execution related items and activities whereas the surface structure layer houses the

dataset that feed the model as well as the supporting infrastructure for providing the dataset. For

instance, IT professionals focus on data acquisition and data access provisioning tasks which

operates within a problem domain of surface level layer. The IT professionals support data

scientists to concentrate their effort in the modeling process where data modeling is in a problem

domain of deep level layer.

Given the abovementioned separation of concerns, a set of discrete predictive analytics

functions can be laid out and automated by organizations as they see fit. Predictive modeling

workflow requires model creation, model execution, model verification and model data to form a

complete predictive modeling lifecycle. A consumer of a PAaaS service can be those who do not

Page 52: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 52

52

possess the modeling skillset (i.e. data mining, model scoring, etc.) but well-versed in business

problem domain (i.e. domain data acquisition and provisioning). This allows business to ignore

details of modeling workflow and deals only with the needed dataset provisioning as shown in

Figure 14.

Figure 14: Automated modeling workflow (Kridel & Dolk, 2013)

Figure 14 describes an automated self-service modeling system workflow reference

model for PAaaS. This approach removes the burden of knowledge from the organization to

implement an all-encompassing modeling workflow from the ground up.

Predictive Analytics Performance Optimization

While the performance in computing system improves and the latency decreases between

connected computing services, it is understandable to expect higher performance when the

modeling services are physically nearer the DBMS where the data were queried. However,

geographical distance imposes a physical limitation and negatively affects network latency

Page 53: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 53

53

between computing nodes which has been a continuing challenge for researchers. Therefore, the

benefits in amalgamating the modeling services and the data services into a single platform

would improve system performance.

This proposition of a single platform approach was described in (Fischer, et al., 2013).

The authors presented a solution in which performing forecasting time series data directly within

DBMS ensures consistency between data and models in addition to a measurable increase in

efficiency. This is largely due to reducing data transfer that made possible by in-database

optimization techniques. As the authors described, there has been a growing trend of integrating

statistical services, predictive and other data analytics into DMBS given the above reasons. For

example, using time series analysis can help finding out the parallel between the forecasting

applications involving sales data, energy data and traffic data.

Time series forecasting has been conventionally used in many of these applications. A

typical time series forecasting process consists of the following consecutive parts: model

identification, model estimation, model usage, model evaluation and model adaptation as shown

in Figure 15. Note that the aforementioned process is cyclical where the final sub-process (i.e.

model adaption) will in turn provide feedback to the model identification and model estimation

sub-processes.

Page 54: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 54

54

Figure 15: General Forecasting Process (Fischer, et al., 2013)

The authors in (Fischer, et al., 2013) pointed out that, although there are evident that the

in-database forecasting process is gaining momentum, in-database forecasting is still lacking in

functionality. The limited forecasting capacities of the major commercial DMBS such as

Microsoft SQL Server and Oracle Database are examples of such concerns. To that end, many

researches have proposed ways to further optimize intra DBMS modeling and commonly break

down into the following items:

1. Integrated Declarative Forecast Queries: Additive keyword such as

FORECAST can be used with a SQL Query statement.

2. Deep Relation Query Processing Integration: Build-in forecast operators in

conjunction with standard SQL operators and seamless data joins with current and

historical records.

3. Transparent and Automatic Query Process: Model attributes can be added

automatically to ad-hoc models for multi-dimensional time series model

processing.

Page 55: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 55

55

4. Query Optimization: Using I/O-conscious skip list data structure for very large

time series data and advance techniques in models reuse in pre-computed multi-

dimensional data.

5. Efficient Update Processing: Model-aware and model-specific mechanisms are

developed to continuously update the underlying model through a stream of

incoming time series values.

To bring to light the current issues of effective model process, the proposed ways include

the abilities to execute forecast queries with optimization built into DBMS. First and foremost,

the suggestion in (Fischer, et al., 2013) is the inclusion of Time Series View alongside with the

traditional view to add support for time series data representation. The Time Series View is

essentially a materialized view with time series attributes incorporated where the date attribute is

treated as first class attribute. It also contains built-in considerations of statistical information

such as standard deviation and confidence intervals. Time Series View combines the support of

continuous data integration where the predicted values from the Time Series View are replaced

with current data as they become available.

Besides the external schema (i.e. Traditional View and Time Series View) discussed

above, a generic forecasting system architecture (Figure 16) based on the traditional

ANSI/SPARC architecture was also proposed in (Fischer, et al., 2013). The architecture involves

a conceptual schema (i.e. model composition) and internal schema (i.e. logical access and

physical access path). This was designed within a traditional relational DBMS environment

containing two separated virtual domains: relation data and forecast models.

Page 56: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 56

56

Figure 16: Extension of 3-layer schema architecture (Fischer, et al., 2013)

The forecasting system architecture as shown in Figure 16 contains the important

components to augment the existing capacities of a relational DBMS. The vertical line in Figure

16 separates the relational model (on the left) and the forecast model (on the right). Each schema

layer in the forecast model incrementally adds native support for the Time Series View. This

design demonstrated the feasibility of extending a relational DBMS architecture to support

forecasting capability without a complete redesign to accomplish a forecast-capable DBMS.

Page 57: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 57

57

CHAPTER III

PREDICTIVE ANALYTICS APPLICATIONS

As discussed in the Explanatory versus Predictive Modeling section of CHAPTER II, we

highlighted the differences between Descriptive Analytics and Predictive Analytics. Descriptive

Analytics is a derivative of all characteristics in an explanatory model, which, has a long history

of research and practical applications that help to describe causality and explanatory statistical

based phenomena. Data Mining has been mostly associated with Descriptive Analytics for

knowledge discovery. Data Mining, in the context of Descriptive Analytics, surfaces the facts

and causal links between variables leading to the discovery of trends and insights. As such, Data

Mining allows one to survey the present situation by pruning away noisy data and ambiguity in a

given complex situation that exhibits a set of convoluted cause-and-effect relationships. These

relationships are often entangled with assumptions, perceptions and potential false information.

While both models have the power to inform, however, only predictive analytics can transform

the practice of decision-making.

The Data Mining techniques have been well researched and understood by researchers

and practitioners. In (Han, Kamber, & Pei, 2011), Han et al. explained in details of the various

Data Mining techniques such as pattern mining, classification, associations, clustering as well as

the advanced techniques involving machine learning methods such as Bayesian belief network

and support vector machines. The aforementioned methods involving supervised machine

learner is particularly suited for Predictive Analytics. One of the many examples is the

Backpropagation artificial neural network. Backpropagation artificial neural network was

invented to solve complex problem in a way that mimic a part of the central nervous systems of a

human brain. The function of a human brain has the ability to learn from experience and

Page 58: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 58

58

recognizes new pattern which is the basis for prediction and decision making. This technique will

be discussed in detail under the Artificial Neural Network section of CHAPTER IV.

To translate the model, methodologies and techniques into practical terms, this chapter

sets the stage for illustration of real-life Predictive Analytics applications across industries. The

discussion thus far has been mostly focusing on the fundamentals of Predictive Analytics. In this

chapter, we will be focusing on the research effort in the practical domain instead of focusing on

the theoretical discussion as we did in CHAPTER II. The varying techniques and methods will

be covered in depth in CHAPTER IV while this section will be focusing on the real-life

applications in Predictive Analytics.

Social Computing

Large-Scale Machine Learning at Twitter Inc.

Twitter, Inc., a well-known online social networking and microblogging service, with

hundreds of millions of active users sending primarily short text messages and search queries,

employ both Descriptive Analytics and Predictive Analytics for sentiment analysis and trend

prediction respectively.

The example of Twitter Inc. touches many aspects of the essay’s subject matter. From a

Big Data perspective, a high volume of messages and events represented data captured through

200 million users at a velocity measured at 400 million daily tweets and a variety of content type

such as text, images and videos (Moore, 2013).

The challenges are evidently shown by the accumulating stream of raw data at a daily

rate of 100 terabyte (Lin & Ryaboy, 2013). To navigate through the mountain of data at petabyte-

scale, the research in (Lin & Ryaboy, 2013) described the Big Data disruption to Twitter Inc. and

Page 59: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 59

59

the novel means to address those problems. The authors identified the problem domains in the

areas of data capturing and data processing.

Firstly, classifying captured data is an important step. There are two classes of data: the

business centric data such as customers and contracts that are part and partial to the business.

This class of data has always been maintained in an organization and it is critical to the day-to-

day business operation. However, there exists a second class of data that represents user

behaviors. This class of data is equally important but often overlooked. This particular kind of

data is crucial to many businesses and especially important to a social media company like

Twitter Inc.

Secondly, the analysis portion of the Big Data equation is the Descriptive Statistics that

includes the use of Online Analytical Processing (OLAP) and Extract-Transform-Load (ETL).

They pervade the majority of many IT operations and they are synonymous to what commonly

known as Business Intelligence (BI). The authors in (Lin & Ryaboy, 2013) recognized the need

to perform Predictive Analytics that cannot be done effectively by traditional methodologies and

tools but it is achievable with advanced modeling techniques such as machine learning.

Thirdly, the open-source disruption, largely credited to the Hadoop open-source

implementation of MapReduce and surrounding technologies that cemented the infrastructure for

Big Data mining.

The contrasting concepts are clear in terms of solving problems that are well-known and

those that are latent. Solving unknown problem requires a new kind of modeling that is designed

for turning vague directives into concrete and solvable problems. The authors also explained that

typical preprocesses of Data Mining, such as data cleansing and data normalization, are still

important for avoiding skewing of predictions as a result of missing data, outlier data and

Page 60: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 60

60

generally incorrect data. The arduous data cleansing process at Twitter Inc. still requires a large

amount of manual intervention and have yet to be fully automated. This is largely due to the

current system complexity as a result of rapid business growth as well as the loose coupling

system architecture design (Lin & Ryaboy, 2013).

Upon the completion of the first step of Data Mining, the authors proceeded to tackle a

business concern of user retention, which is, a stochastic classification problem of what

attributes and behaviors of an active user can be used to predict the future activity of another

given user. That is, the probability of a user becoming inactive is predicted through the shared

attributes of other users who previously became inactive. This problem can be represented as

𝑃(𝑌|𝑋) where 𝑋 is a given user whose inactivity was previously recorded in a set of attributes

represented in 𝑋 = {𝑥0, 𝑥1, 𝑥2, … } and 𝑌 = {𝑦0, 𝑦1, 𝑦2, … } is a set of user attributes and

behaviors for a predictable case of a user 𝑌.

Once a predictive model is developed, the predictive accuracy of a given model can be

assessed through the use of backtesting method. In backtesting, historical records are used in

both model training as well as in model validation. Suppose the entire dataset consists of five

years’ worth of user activity data, the data in the first four years can be partitioned off into a

model training dataset and to be used to train the model. Once the model has been trained, the

model then generates predictions on future user activity. The result of the prediction can be

validated and assessed by the data in the final year of the historical dataset. Thus, backtesting

allows a full assessment of the predictive power of a given model based only on historical data.

Backtesting is achievable because both variables, the prediction outcome and the

prediction criteria are both known facts in the chosen dataset. This allows researchers to compare

the known result (i.e. today’s data) and the predicted result (i.e. computed value) in order to

Page 61: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 61

61

tweak the model parameters to improve overall predictive power. The backtesting method was

the basic form of model validation that Twitter Inc. currently employs.

Outside the internal concerns of Twitter Inc. business operation, the public Twitter.com

API provides a means to externalize some of the massive internal data. Many researchers have

conducted experiments on Twitter Inc. data for various novel measures of hypotheses. The

machine learning approach that underpins the Twitter Inc. platform was discussed in (Lin &

Kolcz, 2012) where integration between the Hadoop Pig platform (Hardoop, 2014) and the

Twitter Inc. massive machine learning engine was discussed in great length.

Another example that leverages Twitter Inc. data was presented in (ARIAS, ARRATIA,

& XURIGUERA, 2013) which described the various ways one can correlate the stock market

volatility, movie box office revenue and presidential polls result by performing sentiment

analysis on Twitter Inc. data. In particular, the daily Twitter Sentiment Index, coupled with the

general sentiment time series data, had been used in a few studies referenced in (ARIAS,

ARRATIA, & XURIGUERA, 2013). These studies showed the correlations between the

sentiment of Twitter users and the forecast targets such as Dow Jones Industrial Index (DJIA)

and US influenza rates.

The result of analysis in (ARIAS, ARRATIA, & XURIGUERA, 2013) was deduced into

a decision classification tree called a summary tree. The summary tree was then built based on

the experimental results with self-tuning capabilities to increase its predictive ability. The

summary tree was not used in the prediction execution process itself but it was used as a

supplementary tool to capture experimental observations. It is important to mention that the

study was conducted in an experimental design where a control group was used to determine

whether the Twitter Sentiment Index alone helped the predictive model to prove the stated

Page 62: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 62

62

hypothesis. The resulting predictions were overwhelmingly positive. The study concluded the

nonlinear models (e.g. support vector machine model and artificial neural network) yielded the

highest successful prediction rate as compared to the result produced by linear models (e.g.

simple regression model).

Like many social networking services, Twitter Inc. suffers from the peril of spam

messages. In (Wang, et al., 2013), the authors employed a Random Tree algorithm for the spam

classification model based on click traffic data and the shorten URLs generated by Bitly.com.

The model produced predictive result with 90.81% accuracy and 0.913 F1 measure value. A fully

developed spam classification model can be used as a predictive model for future spam

detection. This is a classic binary classification problem of 𝑌 = 𝑓(𝑋) 𝑤ℎ𝑒𝑟𝑒 𝑋 =

{𝑥0, 𝑥1, 𝑥2, … }. The model was constructed to classify an email message as either a spam or non-

spam message by predicting a binary target variable of 𝑌. The study showcased the practical use

of classification model to solve tangible business problem of spam traffics in a predictable

manner.

The above discussed Twitter.com case studies underline the convergence of Big Data,

Data Mining and Predictive Analytics in a practical business domain. They exemplified the

application of predictive modeling techniques in solving both known and unknown business

problems.

Network Relationship

In many predictive applications, using only one predictor (e.g. age) to predict a particular

outcome (e.g. education level) is an inexact measure that is problematic due to the lack of

consideration of other important variables. Adding other attributes (e.g. race, height, income

level, etc.) of an entity can improve the validity and accuracy of a measure but the model would

Page 63: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 63

63

still lacks breadth. The reason is that many relationship data between the entities themselves

were not accounted for and very often, they are very important in improving predictability.

The social element in network connections is one aspect where it can enhance data

quality, which is to say, taking into account a person’s social role reveals more data about the

individual’s behaviors. These data would otherwise be hidden had the network connections not

been considered. In (Nankani & Simoff), the authors incorporated the network relations

component in their Classification and Regression Tree (CART) predictive model to enrich the

entity dataset. The authors took into account the network relationships between actors such as co-

authorship, co-participant in a given academic domain to predict and forecast the two outcomes

below:

1. A given research project will be funded or not.

2. The predominant category of personal publication.

As it was noted, the target variables are both a binary label representing funding decision

and a discrete label that defines the publication categories such as book, conference paper and

journal articles.

The methodology depends on the measure of actor centrality in graph theory. Actor

centrality is often used in network analysis which refers to the degree (i.e. indegree or

outdegree), closeness and betweeness of actors (i.e. student, instructor, administrative staff, etc).

By incorporating the network relationship data and by using the Salford systems CART tool

(Nankani & Simoff), the authors were able to enrich existing data to improve the predictive

power of their predictive model that leads to an overall enhanced predictive model. Proving that

the hypothesis that information about network structure can improve the predictive accuracy of a

given model.

Page 64: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 64

64

Education Support

Predicting At-Risk Students

In the realm of pedagogical analytics, maximizing student participation rate often fosters

a positive education environment which is the goal of many educators. In (Annika Wolff, 2013),

the authors presented an interesting correlation between clicking behavior of students and the

course outcome of students in an Virtual Learning Environment (VLE) conducted at The Open

University Institute. The result of the analytics work was used to predict at-risk students so

interventions can be administered in advance.

Generally speaking, the VLE activity information (e.g. hyperlinks click frequency) was

considered a strong predictor in many pedagogical researches in an online learning environment.

This is because, a number of problematic obstacles exist exclusively in distance learning

environment that do not exist in canonical educational institutions such as geographic and time

zone differences, lack of in-person face-to-face learning environment, imbalance of educational

background among students. The authors employed a number of independent variables to predict

the dependent variable of at-risk students whom tend to either fail a particular course or dropping

out a program entirely.

Besides the activity data in VLE, the authors also incorporated financing, demographic,

course subject area and general course information as independent variables in the study. The

authors pointed out the danger of misinterpreted correlation such as click frequency, which

represent only an aspect of student activity measure, does not always correlate with the course

outcome. The perception of students with high click frequency leads one to presume an engaged

student in terms of online access frequency. Conversely, a low click frequency would imply a

low engagement student and consequently lead to a negative course outcome. In reality, student

Page 65: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 65

65

preference on learning material should also be taken into account. Preference is a latent factor

that cannot be measured by activity data as some students prefer printed materials over online

materials and therefore it would skew the abovementioned model result.

The subsequence steps demand the consideration of outcome classifiers which are the

classification labels for what outcomes constitute an at-risk student. The authors defined such

labels as performance drop and the binary response of course outcome (i.e. pass or fail). The

execution of model was done in a multiple combination of predictor variables, such as,

standalone VLE Activity data, standalone Tutor Marked Assessments (TMA) data, VLE and

TMA data, etc. The resulting predictor performance on performance drop outcome favors a

combined data categories (i.e. VLE and TMA) based on decision tree algorithm. The course

outcome dependent variable (i.e. pass or fail) showed a different picture where VLE only data

(i.e. total clicks, delta in clicks, clicks relative to historical context, etc.) yielded the best

predictive performance based on the precision, recall and f-measure scores.

The other permutations of the model involved the inclusion of demographic data which

had been confirmed to improve overall predictive power. Also, over the course of the assessment,

the predictors performed in varying degree of accuracy and precision at each point in time.

Exclusive VLE data yielded better score during the early stage while other forms of measures

produced better overall result at the later stage of the experiment.

The authors concluded the paper with a confirmed hypothesis based on the above

predictive model that, click frequency in online learning environment do correlate positively with

course outcome as long as the measure takes into account the overall timeline and clicks

volatility. For instance, if click frequency of a student decreases abruptly during a course, this

Page 66: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 66

66

correlates positively to a negative course outcome. Postulating such hypotheses and confirming

them by analyzing and predicting student performance in a VLE is important to many educators.

Predict Course Success

In (Barber & Sharkey, 2012), the authors conducted an empirical experiment at the

University of Phoenix where multiple predictive models were created based on a dataset

containing 340,000 student records. The authors proposed three models with each containing

different independent variables as predictors for course success prediction.

Model 1 is a Logistic Regression model targeting an ordinal dependent risk level variable,

the values are, high, neutral and low risk. While model 1 performed admirably at 90% accuracy

on pass (i.e. low risk) and fail (i.e. high risk) with 𝑝 < 0.5, the authors expressed concerns over

the model. The data that supported model 1 originated from multiple databases with varying

level of data quality that required significant data cleaning and perpetration in order to bring the

data to an acceptable level. Also, the researchers had reservation over the level of granularity of

the risk level with the concern of levels being over generalized.

Model 2 was built using Naïve Bayes algorithm in RapidMiner and it included variables

that were not present in model 1 such as military status and financial status as well as the

inclusion of the discussion board posting count. In model 2, the researchers omitted a few

independent variables that are insignificant and lack predictive power through previous

observations in model 1. These include the attributes of gender and age, amongst others. Model 2

showed improvement over model 1 for the weekly predictive accuracy measure reaching 95%

accuracy by week 3. The most influential independent variables in model 2 are credits earned

versus credits attempted ratio, previous financial status and the most powerful predictor being

cumulative points earned.

Page 67: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 67

67

Model 3 was built with the hypothesis that student engagement is positively correlated

with course pass grade; this is in alignment with the hypothesis discussed in the previous

Predicting At-Risk Students section. The independent variables under consideration included a

lower form of discussion board activity data that could suggest engagement level (i.e. time since

last course, public post versus private post to instructor).

The authors concluded the paper with the emphasis on model accuracy and utility which

are the two topmost considerations. Utility refers to whether the resulting model result is

actionable and can yield fruitful outcome for student success. The experiment also highlighted

that the combination of differing predictor variables with the varying predictive models, would

yield a very different prediction outcome.

Video Gaming

The video gaming industry has long been engaged with machine learning. The video

gaming industry is one of the early adopters in Artificial Intelligence (AI) research and

development as far back as 1959 (Kaur, et al., 2013). Video gaming industry, also called

interactive entertainment industry, is intrinsically tied to the feedback mechanism of the

anticipation of human responses by establishing mutual and reciprocal human-machine

relationship. An enjoyable video generates human emotional responses which is a direct result of

machine learning techniques. Some of these emotional responses are amusement, excitement,

contentment, wonderment and surprise.

In (Geisler, 2002), a great amount of details went into explaining enemy opponent AI

design in First Person Shooter (FPS) video game. The feature vector representing player

decisions (i.e. accelerates, changes movement, changes facing, or jumps actions) are captured for

supervised learning through the learning algorithms of ID3 Decision Tree, Naïve Bayes and

Page 68: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 68

68

Artificial Neural Network. The learning algorithms reinforce the game’s AI agents such as emery

opponents.

More recently, researchers have devised techniques to create supervised machine learning

models by capturing human gameplay activities and behaviors through video game to create a

human-like AI model. This is an effective way of using human actions to build model that

embodies human behaviors.

One such example is to use machine learning models to solve real-world problems

through video gaming as explained in (Sanfilippo, et al., 2011). The authors proposed and

developed a prototype system for generating predictive outcome in assessing the propensity for

state and non-state actors in participating illicit nuclear trafficking through Bayesian belief

network and agent-based simulations. The goal of the model design proposed by the authors was

to capture real-life human interaction behaviors through video game simulation. The data are

then analyzed through machine learning techniques based on the Technosocial Predictive

Analytics (TPA) framework. The application for the model in (Sanfilippo, et al., 2011) has a very

strong focus on the deployment of Predictive Analytics in the field of government and military

surveillance, which was used to predict illicit activities for the purpose of resource allocation and

strategic resource deployment to high risk regions.

The simulation itself was constructed as a multiplayer game where players took the roles

of actors in a staged illicit trafficking scenario based on the following conceptual components:

Technosocial Modelling, Knowledge Encapsulation Framework and Analytical Gaming. The

authors described the aforementioned components in great depth but the main thrust of the

concept as it relates to Predictive Analytics, was the inclusion of System Dynamics and Bayesian

Page 69: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 69

69

Belief Network methods through supervised learning. These methods are supported by a training

dataset that was created by multiplayer gaming sessions simulating real world events.

The Bayesian Network model as shown in Figure 17 was selected in modeling the illicit

trafficking game experiment as the learning model. This is because the stochastic nature of

player reactions to random events can lead to a number of possibilities that are non-deterministic

in nature. Also, System Dynamics modeling were used to map model parameters (e.g. intent to

establish alliance) to game parameters (e.g. initiate communication) to control elements of game

environment. Using Bayesian Network and System Dynamics are a two-prong approach in

model selection to tailor specialized aspects of the game engine.

Figure 17: Bayesian Net Model of Intent to Proliferate (Sanfilippo, et al., 2011)

Creating strategic and tactical human responses is a difficult proposition for unsupervised

machine learning techniques, especially, given the fact that human contextual information is

Page 70: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 70

70

acquired through experience. Contextual information is a necessity in devising strategic and

tactical decisions. Information arrive at human through a number of random and established

channels. Structure channels such as television news is very much different from other

unstructured information sources such as social interaction or word-of-mouth communication.

Particularly, in offline communications involving verbal and non-verbal information are not

commonly captured to be analyzed by machines. Thus, these types of contextual information

often exist solely in human brains. To that end, a paper by (Riensche & Whitney, 2012) fused the

two concepts of predictive analytics and gaming succinctly by proposing knowledge transfer

from human to model through gaming. This maximizes the information input to machine

including contextual information. The paper focused on war game simulation which is type of

Analytical Gaming, a form of Serious Gaming that dealt with players’ ability and facilitates

knowledge extraction.

Law Enforcement

In criminology, the study of criminal behavior of individuals as relates to social science is

an interdisciplinary field that deals with, at the very least, behavioral science and law

enforcement theory. Many results of the criminology study have been directly influencing

lawmaking as well as legislations of the criminal justice systems. The use of quantitative

statistical methods in criminology in identifying chronic offenders have long been used and the

practice is dated back to 1972 (Jennings & M.C.J, 2006) in order to predict high-risk offenders.

The predictive model led to the construction of prediction instruments for measuring

criminal behavioral risks. In the article by (Jennings & M.C.J, 2006), the authors discussed the

practice of criminal classification by means of risk assessment instruments. The risk assessment

instruments draw on data from a wide area such as psychological data, socioeconomic data,

Page 71: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 71

71

demographic characteristics and conditions. The result of such measure had led to the

intelligence-led policing such as proactive policing, problem-oriented policing, community-

based policing and knowledge-based policing. They are the constituents of an overarching

forward-looking model of crime prevention that relies on risk assessment techniques for both

new offenders and reoffenders.

The paper by (Jennings & M.C.J, 2006) discussed the result of an empirical experiment

with a randomized sample of offenders whom have a history of arrests within the past three

years. The sampled offenders were studied over the course of six months and the observed

behaviors were recorded including the number of re-arrests made during the six months period.

The independent variables for the prediction model were selected as following: property crimes

scale, person crimes scale, drug/alcohol scale, crime severity scale, repeat offender scale and the

violence scale. The dependent variable in this case was the recidivism rates of a prior offender.

The result was calculated using principal components analysis method and the result

suggested a strong indication of a positive correlation amongst the aforementioned six factors.

The final result from the chi-square tests led to the following table of figures for the three risk

levels of recidivism, displaying a strong positive correlation between risk levels and classes as

shown in Figure 18.

Page 72: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 72

72

Figure 18: Group Recidivism Rates (Jennings & M.C.J, 2006). Note: n=number of individual per category (scale) and per risk

class, %=percentage of the individual in a specific class were re-arrested.

Before predictive analytics became a recognized term, many researchers in criminology

had used the same statistical approach in identifying high risk individuals with the intent to

balance individual rights and public safety, particularly in the area of reoffending and recidivism

risk assessments. According to the article by (Greengard, 2012), a growing trend amongst law

enforcement and modern policing is to use data and model driven mathematical and statistical

techniques to better direct crime prevention. Like other predictive modeling applications,

predictive policing makes use of high dimensionality dataset spanning across space (e.g.

locations) and time (e.g. time of day). They are made up of past historical crime records that can

be deduced into geographical, temporal and distribution correlated patterns.

Human behaviors often carry a certain pattern in a form of habit which applies to

criminal activities as well. For instance, it is statistically proven that the probability of a past

crime event would increase the likelihood of higher future crime activities occurring within the

same vicinity of a past crime. This can lead to a concentrated occurrence of crimes due to a

collective social presumption of tolerable activities. This phenomenon is called a Broken

Windows Effect. The effect can be observed through the forming of crime hotspot clusters where

a prolonged deterrent action has not been taken in those areas with high criminal activities.

Page 73: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 73

73

Generally speaking, law enforcement has policies in place to review reports of past

crimes stored in databases for the purpose of crime mapping. The move from reactive policing to

proactive policing is not a new trend in law enforcement practices, but the shift to a stronger

reliance on proactive and predictive approach is in fact gaining momentum. The proactive

policing translates into patrols deployment in areas that require crime deterrent. One of the many

methods in proactive policing can be done through crime pattern analysis.

Predictive policing, on the other hand, requires advanced analytical tools and techniques

to increase crime predictability for the required accuracy. One of the key examples in predictive

policing was demonstrated by the Memphis Police Department’s intelligent crime fighting

solution, called the Blue CRUSH (Criminal Reduction Utilizing Statistical History). The design

goal of Blue CRUSH is to surface actionable insights from two commercial analysis support

solutions: IBM SPSS Statistics and ESRI ArcGIS. Blue CRUSH resulted in a 30% reduction in

overall crimes and 15% reduction in violent crime in the four-years span and directly responsible

for 50 arrests on drug related crimes since the Blue CRUSH deployment.

Another example of predictive policing application was presented in (Hollywood, Smith,

Price, McInnis, & Perry, 2012) by the National Law Enforcement and Corrections Technology

Center. The authors were clear on correcting the perceptions held by many people in regards to

the effectiveness of predictive policing. The authors stated that predictive policing does not

equate to divination and that we should perceive predictive policing as a decision support model.

The long practice of crime mapping is considered a basic form of predictive policing to

anticipate criminal activity. Crime mapping is an assistive and informative tool for crime

prevention. The predictive policing process mirrors many iterative-based processes where

forming a feedback loop is necessary. The feedback loop allows the result of a single process

Page 74: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 74

74

iteration to enrich the input of subsequent iterations. The predictive policing process

(Hollywood, Smith, Price, McInnis, & Perry, 2012) comprises of four high level sub processes:

data collection, data analysis, police operations and criminal response.

In fact, many crime hotspot identification methods involve a combination of multiple

techniques such as near repeat methods, risk terrain modeling, regression and other common data

mining methods as well as spatiotemporal methods. These mathematical methods have been

studied in great details in academia as presented in (Short, D’Orsogna, Brantingham, & Tita,

2009). They are not exclusive to predictive policing but the methods apply to Predictive

Analytics in the most fundamental ways. These methods allow the deduction of past crime data

to produce insights that represent crime patterns. Many of which are very often temporal and

spatiotemporal relative, in terms of the day/night cycles, weekend versus weekday, paydays,

sporting events, concert events and time of year. These independent variables are used in

correlation analysis to help predict future crime activities.

The potential of predictive policing is worthy of continuous research within the realm of

Predictive Analytics. An assessment was conducted in (Yang, Wong, & Coid, 2010) to assess the

efficacy of violence prediction led by predictive policing. The authors highlighted the difference

between high-frequency and low-frequency crimes where low-frequency crimes such as serial

killing and school shooting tend to generate many false positive type I errors. In those cases,

Type I errors are costly and socially prohibitive to act upon.

The simple model of assessment-to-prediction-to-intervention is a common model for

actuarial risk assessment and Predictive Analytics alike. The actuarial risk assessment model is

an accepted standard of forensic risk assessment practice which draws in multiple constructs

ranging from clinical (e.g. personality disorder) to situational (e.g. community support). In

Page 75: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 75

75

actuary, the selection of varying constructs changes based on the subjects of assessment. The

practice of actuary derives predictions based on empirical evident and professional judgment.

The method is subjected to repeated empirical validation while keeping both static (e.g.

ethnicity) and dynamic (e.g. received treatment) predictors under consideration.

The authors discussed in depth of the relationships established on violence and

psychopathy with respect to Predictive Analytics. The comparison of predictive efficacy was

performed on 9 actuarial assessment instruments: Hare Psychopathy Checklist-Revised (PCL-R),

Violence Risk Appraisal Guide (VRAG), Violence Risk Assessment Scheme (HCR-20), Level of

Service Inventory-Revised (LSI-R), Psychopathy Check List: Screening Version (PCL:SV),

Lifestyle Criminality Screening Form (LCSF), General Statistical Information on Recidivism

(GSIR), Sexual Violence Risk-20 (SVR-20) and Static 99. The predictive efficacy of

abovementioned instruments were studied and concluded with similar performance given the

same context. The majority of the variance in efficacy was due to methodological features such

as age and length of follow-up, which is to say, the conclusion was such that none of the

instruments studied can produce significant standalone advantage that are able to differentiate

itself from the other instruments in terms of predictive power, given the same methodological

features.

The result suggested that the efficacy of the simple tools (e.g. summing) might have

reached a plateau and the needs for novel means of identifying and combining risk predictors is

in order. This serves as a great remainder to consider the significance of the role of data in

Predictive Analytics.

Page 76: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 76

76

Business Applications

Many business applications of descriptive analytics are customer centric, meaning that

the goal of many businesses is to improve customer service experience in exchange of business

loyalty. Companies add tangible business values to their businesses by having a better

understanding of their target customers, leading to an increased demand of analytics that aids

decision makers to better drive business decisions to thrive amongst competitors. Traditional

Business Intelligence (BI) answers standard queries that are descriptive analytics driven, which

is to say, queries that answer only in historical context based on descriptive and explanatory

models. The models are designed to work with past dataset. While predictive analytics also

depends on historical data, and that data is in fact historical by definition, the key difference is

the timeliness of the data and the type of model used.

In (Nauck, Ruta, Spott, & Azvine, 2006), the authors explained the common approaches

for data analysis and highlighted the inefficiency in many linear and nonlinear regression based

modeling methods. The regression model often lacks depth in terms of revealing

multidimensional variables dependencies and it is not designed to reveal latent variables while

factor analysis excels at that. The authors proceeded to base their analysis in Bayesian network,

citing a superior alternative to regression based analysis. Bayesian network, based on Bayesian

Theorem, is a way to discover information about the structure of statistical dependencies. The

authors employed a Bayesian networks modeling tool called Intelligent Customer Satisfaction

Analysis Tool (iCSat), to perform overall sensitivity analysis and what-if analysis in supporting

customer satisfaction analysis. The goal of the analysis was to identify customers whom are in

jeopardy (e.g. potential high churn rate), customer satisfaction target setting and field force

performance.

Page 77: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 77

77

Bayesian network deals with classification problem based on Bayes’ Theorem. Given a

class label 𝑌 and the dataset 𝑋, the probability of 𝑌 occurrences in 𝑋 is determined by 𝑃(𝑌|𝑋).

That is, the probability of 𝑌 happens given 𝑋. The resulting formula for a single condition is

determined by the formula shown in Figure 19.

𝑃(𝑌|𝑋) =𝑃(𝑋|𝑌)𝑃(𝑌)

𝑃(𝑋)

Figure 19: Bayes' theorem formula

When dealing with multiple conditions, the results of the formula (Figure 19) are

multiplied per each condition to produce the overall probability value. Bayesian Belief Network

can be seen as a way to visualize a collection of naive Bayesian rules and allows for a way to

determine probability values based on the combined relational structure between dependent and

independent variables. The dependent and independent variables are also directionally

significant. A non-bias and a priori way to calculate probability based on a nested set of

variables. This concept is deeply rooted in Bayes’ theorem.

Given the complexity involved in a large size Bayesian Network model, the authors

relied on iCSat to build models and to perform data analysis that deals with what-if scenarios.

This is possible because of the nature of structural relationships of a directed acyclic graph.

Bayesian Network can handle noise and partial information using local and distributed

algorithms for inference and learning (Pearl & Russell, 2000).

Another example of Business Application was illustrated in (Waller & Fawcett, 2013),

the authors discussed the key concepts where the exploitation of domain knowledge, as evidently

expressed in the variety of skillsets required by predictive analytics research, including statistics,

forecasting, optimization, discrete event simulation, applied probability, data mining and

analytical mathematical modeling, are a reflection of market demand for specialized skillsets to

Page 78: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 78

78

solve problems within a specific domain. The predictive power of predictive analytics in

combination of Big Data augments the field of Supply Chain Management. Optimizing logistics

with respects to customer, sales, carrier, manufacturer, retailer, inventory, location and time, is

curial and it is key to balancing cost and service level, which are of primal concern to many

businesses.

The result from business optimization through Predictive Analytics directly adds values

to businesses and optimization necessitates the higher volume, variety and velocity properties of

Big Data as the foundation for analytics. Predictive Analytics helps researchers and practitioners

to analyze the multifaceted field of Supply Chain Management in that it approximates the

relationships between variables with the use of statistical and mathematical deduction methods,

to draw predictions based on the result from data mining historical data in both quantitative and

qualitative means.

Financial Engineering

The survey discussion thus far has identified a number of applications where Predictive

Analytics can directly or indirectly influences the mode of established operations to further

enhance existing capacities and horizontally scale across industries and organizations. It is

evident that Predictive Analytics possess a crosscutting solution that deals with issues and

challenges faced by many academic institutions, business corporations and organizations with

specialized functions. However, none other has generated the same level of controversial public

discourse that took place in financial service sector. The emphasis has been on finance

investment where financial engineering is a strong focus. The field of financial engineering is

very much grounded in the notion of prediction.

Page 79: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 79

79

The machine learning application in investment finance is prevalent in the industry. This

is because the multifactor market performance variables are vital to investment decisions and

business strategy. Machine learning methods used in predicting security price patterns are a

prominent example of Predictive Analytics for stock traders and portfolio managers. This view is

evidently supported in (Soulas & Shasha, 2013). The authors employed the Statstream software

to correlate securities using a sliding window technique and to predict foreign exchange rate

changes.

In financial investment services sector, the institutional investment firms such as hedge

fund companies, pension fund companies, mutual fund companies and insurance businesses,

collectively known as the buy-side firms, are innately dependent on prediction. They often make

investment decisions based on recommendations from the sell-side firms such as market research

companies and brokerage firms.

Traditionally, the data input coming from the sell-side firms arrive through casual

channels such as conversations that occur during meeting and between phone calls. Quantifying

information that comes from casual conversations is problematic as the information obtained

does not get propagated throughout the rest of the system. This reduces the chance to capture

information that lead to a traceable and actionable trading recommendation. Many institutional

investment firms have developed in-house algorithmic trading platform to automate trading

executions as well as for trading decision support. These algorithmic trading systems often rely

on a constant stream of data from multiple sources in order to derive trading decisions that leads

to a binary decision of buy or sell of a particular security.

Hedge fund companies in particular, have developed such proprietary algorithmic trading

systems to provide Predictive Analytics functions. To that end, many trading algorithms sift

Page 80: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 80

80

through massive volume of market information based on predictive modeling to reach at a

prospective decision based on the potential performance of any given security and market

conditions. Rapid data feed come into the system that varies in scope and in kind, characteristics

that are shared by Big Data.

In many ways, the market information resemble the definition of Big Data of

unstructured dataset where variety means data from blogs, published news, interviews, phone

conversations and other forms of unstructured communication that are in high velocity and

volume. They fused together with advanced modeling techniques involving the calculation of

alpha (α) and beta (β) coefficients based on regression model (e.g. Capital Asset Pricing Model)

to access risk level on expected return. Therefore, the application of algorithmic trading is the

prime example of integrating Predictive Analytics in Data Mining with Big Data.

One such system is called Alpha Capture System (ACS) as discussed in (Thomas, 2011).

ACS is a collaborative and quantitative research platform for buy-side firms to track, rank and

audit as well as to reward sell-side firms for their investment analysis idea inputs and trade

recommendations. The ACS provides a single point of contact for the submission of trade ideas

which came from semi-structure data format that combine a binary recommendation (i.e. buy or

sell) and the textual analysis report. Prior to downstream processing, the data will be cleansed

and data-mined and subjected to ideas classification and prioritization based on securities

attributes such as firm type and investment dollar amount.

Marshall Wace hedge fund is one of the first fund investment firms to design and develop

a homegrown Alpha Capture and Generation System (ACGS) called Trade Optimized Portfolio

System (TOPS) as early as 2001 (Thomas, 2011). TOPS ranks trade ideas offered by sell-side

researchers through Data Mining and makes trade decisions algorithmically based on Predictive

Page 81: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 81

81

Analytics. These practices of ACS have garnered endorsement from the Financial Services

Authority (FSA) of UK for its traceability, auditability and build-in accountability for both

reward and penalty. The decisions from ACS are transparent in terms of system dataflow. In the

report by (Financial Services Authority (FSA) UK, 2006), the authors clearly stated that ACS

promotes good practices that are policy driven, risk assessed, compliance and alert capable, audit

focus and generally considered a transparent system. The FSA regards such system a major step

forward that can avoid insider abuse and other criminal offences as compared to the traditional

means of communications.

In this section, we highlighted the integration of Big Data and Predictive Analytics in the

most practical sense through the illustration of financial engineering. The background of

financial engineering exemplified Big Data in which it unifies social media data and expert

opinions to derive automated trading decisions. All of which exploit the massive volume, rapid

velocity and ranging variety properties of Big Data.

Summary

The practice of Data Mining had undergone a tremendous shift in adoption and

application in both the academia setting and business setting. As pointed out by (Piatetsky-

Shapiro, 2007), the value of applying Data Mining in many industries, had been realized and

cemented in our collective understanding. The meta-analysis approach of this essay is to contrast

different studies across many industries done by researchers and practitioners involved in

actuary, forecasting, statistical modeling, computer science, machine learning and business

intelligence. These are only a subset of a small sample of the overarching reach that Predictive

Analytics can extend to.

Page 82: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 82

82

There are evident that shows the pervasiveness of Predictive Analytics in our society. As

a researcher, this essay has illustrated a number of research articles related to the subject matter,

as well as discussion in a variety of applications. As a consumer, online recommender systems by

companies such as Amazon and Netflix are prime examples of our direct reliance on Predictive

Analytics. As a citizen, predictive policing techniques employed by law enforcement agencies

directly improve the safety of the people they serve. As a patient, pharmaceutical companies rely

on Predictive Analytics to assist drug research and development as well as genetic research,

which directly affect us in terms of medical advancement and disease control. The examples are

numerous. The effects of Predictive Analytics have made significant impacts on our lives,

sometimes surreptitiously, including a lesser known application in bioterrorism surveillance

(Berndt, Bhat, Fisher, Hevner, & Studnicki, 2004). The industries of social computing, general

business applications, education and pedagogy support, video gaming, law enforcement and

financial engineering are just some of the applications discussed in this essay. Evidently, at a

minimum, Predictive Analytics can enrich existing decision support models given the discussion

in this chapter. The effects of Predictive Analytics go beyond the applications and industries that

we have discussed thus far.

To conclude, we discussed the twitter case study and concluded that marrying Big Data

with Data Mining to transform machine learning techniques into measurable and actionable

predictions for spam classification using cutting-edge infrastructure and open source toolset.

Also, we discussed how Predictive Analytics can support educators in helping students early-on

in the learning process by predicting at-risk students whom show signs of distress and negative

behaviors correlated with course dropout. We discussed video gaming industry, which is, one of

the early adopters in machine learning research with the goal to create highly human-like

Page 83: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 83

83

behaviors of Artificial Intelligence agent design. Predictive Analytics helps game designers to

make video games interesting and engaging to appeal to video gamers by evoking human

emotions. In areas where Predictive Analytics can be a lifesaving supplement, law enforcement

continues to use Predictive Analytics to identify crime hotspots with systems that alert police for

violent crimes prevention. Health industry applies Predicative Analytics on biometrics

information to help physician to monitor life threatening vital signs.

The aforementioned industries capitalize on the methodologies and methods of Predictive

Analytics to extract the telltale signs of the solutions to the problems they are to solve. These

signs are often buried deep within their data. The facts are often captured but scattered across

multiple domains in a form of distributed dataset, calling for Data Mining methods to reconcile

the knowledge embedded within.

Page 84: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 84

84

CHAPTER IV

METHODOLOGY

The format of this essay took a qualitative approach of meta-analysis to study and survey

the current landscape of Predictive Analytics in Data Mining with Big Data. As such, the

methodology employed in the creation of this essay is neither observational nor experimental.

The development of this essay relied greatly on previously peer-reviewed literatures. They are to

allow inferences made from formerly conducted observational or experimental exercises to

support the arguments made herein. The experiences derived from various online research outlets

such as IEEE, ACM and SpringerLink are the main sources of information in answering the

proposed questions stated in CHAPTER I.

The approach to the research subject is done qualitatively with a specific focus on the

subject matter based on grounded theory, so the approach taken for this research is bottom-up

and inductive in nature, through the ideas and information embodied by the literatures reviewed.

As such, in alignment with the grounded theory principles, which are to help us to understand

complex problems through a comprehensive, systematic and inductive approach to developing

theory. This research does not provide a hypothesis but attempts to generate a theory for the

research constructed through existing research results published by other researchers. The

research result of this paper also includes a set of taxonomy diagrams (shown in Figure 1, Figure

3 and Figure 6), identifying the terms, methodologies, techniques and methods used in Predictive

Analytics as it relates to Data Mining. This is in alignment with the research objective to survey

the technological landscape of the subject matter.

The questions stated under the Research Problem section define the limits of the scope of

discussion in order to avoid scope creep. While the proposed problems are not definitive, as they

Page 85: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 85

85

are meant to set a tone for the paper, they are meant to take the first step in reconciling some of

the overlaps and misunderstanding between the well-known disciplines and practices. Also, they

define the categories of research discussion as a way to delimit the research periphery in the

presence of voluminous research papers available on the subject matter. The research process

does not only rely on academic papers but also other forms of publications such as technical

articles, information technology magazines, conference proceedings, textbooks and other online

materials that would help the research effort. However, the focus of source information

references and citations come from peer-reviewed academic papers obtainable through reputable

outlets such as IEEE and ACM online databases to ascertain information trustworthiness.

The diverse sources of information helped to develop the research constructs and the

progressive improvement in attaining construct validity. The multiple information sources

provide the basis for triangulation of facts, statistics and events that are relatable and traceable

through previously peer-reviewed observational and experimental studies. Since the purpose of

the research paper is to survey the information technology landscape, conducting an

experimental research is not required. As such, aspects of the research will be quantifiable

through online datasets and the result generated from previous research studies.

The primary research tool is a personal computer with Windows 8.1 as the operating

system. The personal computer has unrestricted internet access and has sufficient hardware

power to support the applications and tools needed to aid in the research. The software and

online services for the research includes, but not limited to, the software and services listed on

APPENDIX B under the following sections: Productivity Software, Internet Browsers, Open

Source Predictive Analytics and Data Mining Tools, Python Related Statistical Libraries,

Page 86: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 86

86

Literature Search Engines, Research Paper Online Databases, Research Management Tools and

Services and Online Communities sections.

Preliminary Discussion

The common theme in the topic of Predictive Analytics has been identified as an

advanced technique to forecast future outcome in both population domain and individual basis.

The majority of the research papers referenced in this essay have utilized Data Mining

techniques where Predictive Analytics was mentioned. This indicated some of the techniques in

Data Mining are treated as prerequisite to Predictive Analytics. Particularly, in the paper by

(Nyce, 2007), stated by the author that, using Data Mining techniques to avoid the “garbage in,

garbage out” modeling conundrum is first and foremost to any kind of data analysis including

Predictive Analytics.

Some of the research papers indicated the use of Big Data in the context of Data Mining,

however Big Data is not prescriptive within the domain of Predictive Analytics in Data Mining.

While certain industries can benefit from Big Data such as insurance industry and finance

industry in credit scoring (Nyce, 2007) and SCM (Waller & Fawcett, 2013) to optimize business

operation and logistics performance, as well as retail business for targeted marketing and

advertising. Not all Data Mining and Predictive Analytics applications mandate the use of Big

Data. However, Big Data do represent a major enabler in certain applications and industries.

Data Mining and Predictive Analytics are inseparable in this discussion within the context

of academia while Big Data is supplementary to the discussion. Big Data adds an additional

dimension to the already complex multidimensional field of Data Mining. As previously

mentioned in CHAPTER I, employing Big Data means uprooting our collective understanding of

the most fundamental database technologies. In that, the commonly employed approach to the

Page 87: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 87

87

management of Big Data is the non-relational NoSQL model, as well as the distributed and

parallel processing in computing model. Regardless, the trade-offs exist in employing Big Data

and it can be complementary in most research subjects and fields of study as indicated in

CHAPTER II and CHAPTER III.

The benefits Big Data brought to the research community are the simple fact that it

provides an additional data dimension that houses the latent variables, which, allows for

uncovering correlations that are normally unobserved. For example, the system behind Google

Flu Trends (Google Inc, n.d.) has the ability to forecast influenza outbreak faster than U.S.

Centers for Disease Control and Prevention (CDC) in previous years (CBC, n.d.). In this

particular case, the two distinct approaches employed by two very different organizations (i.e.

Google and CDC) show prediction accuracy favoring those that include Big Data (i.e. Google). A

recent study by (Preis, Moat, & Stanley) predicted financial market volatility using Google

Trends and produced accurate market performance predictions. The underlying premises of

Predictive Analytics are correlation and causation of factors, incorporating Big Data helps to

broaden our view of research and analysis to account for unseen constructs. Thus, Big Data can

enhance the predictive power of Predictive Analytics provided that we have accurate and reliable

data through Data Mining and well defined models through statistical modeling.

Another observation made in Predictive Analytics is that it is often temporal dependent.

For instance, the paper by (Maciejewski, et al., 2011) linked spatial and temporal views to

analysis and forecasting which led to the use of time series analysis and prediction as a means to

discover hotspots. The aforementioned examples of Google Trends are also temporal dependent.

Time is one factor that exists as an independent variable in almost all research papers

encountered thus far. In fact, many research methods such as cross-sectional study and

Page 88: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 88

88

longitudinal study are inherently time and location dependent. As such, putting data in the

context of time is an important aspect of ensuring research accuracy and precision.

Preliminary Analysis of the Topic Data Using Google Trends

The data collection for this essay is primarily done through literature review. Google

Trends was used to uncover some of the interesting phenomena. As a preliminary exercise, one

of the approaches is to correlate the point in time with the search terms where there had been an

increase in interests of the subject matter. The three search terms (i.e. Data Mining, Predictive

Analytics and Big Data) were submitted to Google Trends as shown in Figure 20, Figure 21,

Figure 22 and Figure 23, with an emphasis given to Figure 23 where a convergence between the

red line (i.e. data mining) and the orange line (i.e. big data) can be observed in year 2013, hinting

to a correlation of interests amongst the three terms. The charts also reflected the history of the

terms used within academia with the term “Data Mining” leading in search volume since 2005.

This is due to the relative early establishment of research interest and adequate understanding in

the concepts of Data Mining dating back to 1996. Note that the highlighted (in yellow) dotted

line represents the predicted trends in volume.

Figure 20: "Predictive Analytics" Search Term - Google Trends Chart, February 2014

Page 89: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 89

89

Figure 21: "Data Mining" Search Term - Google Trends Chart, February 2014

Figure 22: "Big Data" Search Term - Google Trends Chart, February 2014

Figure 23: Combined Search Terms - Google Trends Chart, February 2014. Blue color = Predictive Analytics, Red Color =

Data Mining, Orange Color = Big Data

To quantify the collected data, the first step is data selection and noise reduction, that is,

remove unnecessary data to produce a higher signal to noise ratio. In the case of this research,

the noise data can be represented by research discussions that are irrelevant to answering the

research questions. Regardless, the step-by-step approach in the entire process spectrum of KDD,

Page 90: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 90

90

from collection to visualization, will be carried out qualitatively and quantitatively using some of

the recommended tools and services as detailed in this essay.

Discussion

The study of information systems technology enables the understanding of how human

interact with computer systems. In the broadest sense, human provide instructions as inputs to

computer systems for execution, the result of the execution are presented in audiovisual

responses. This encapsulates the basic form of human-machine communication. Unlike human-

to-human communication, human-to-machine communication lacks richness in context. Human

communicates with each other in ways beyond the spoken and written language. Human

communication includes contextual rich information such as body language (e.g. hand gesture

and facial expressions) and entity relationship (e.g. friendship, kinship and acquaintance).

Human-machine communication, however, have yet to reach the fullness and richness of

information exchange that inter-human communication provides.

Context-Awareness and Big Data

Many researches have been done to expand on the idea of context-aware computing and

indicated that the definition of the term “context” changes meanings overtime. The definition of

the term “context” changes parallel to the evolution of computing systems. The definition also

changes parallel to the increasing reliance human has on machines for data analysis. From the

location-centric definition in 1992 to the history of previous interaction in 2008 and the 2012

definition focus on context discovery (Kiseleva, 2013), researchers have continuously

supplement computer systems with contextual information in hope to increase machine

intelligence as it relates to human cognition and communication. The goal is to increase the

overall analytical power of computer information processing by incorporating contextual data

Page 91: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 91

91

that are dynamic and relevant in any given situation. The evolution of context definition between

1992 and 2012 is shown in Figure 24.

Figure 24: The evolution of context definition (Kiseleva, 2013)

The following definition of context-aware computing was given by (Dey, 2001):

“Context is any information that can be used to characterize the situation of an entity. An

entity is a person, place, or object that is considered relevant to the interaction between a user

and an application, including the user and applications themselves.”

The above definition still hold true today in the face of the dynamic nature of what we

consider as contextual information. At a high level, contextual information is entity-associated

situational information that is relevant to the interaction between user and application. The

example of an entity is clear, a person (e.g. user), place (e.g. location) and object (e.g. mobile

device). However, what constitute context remains an abstract definition as it is dynamic in

nature. It is important to recognize contextual information as variables that exist outside the

normative independent variables for a given analysis. For instance, spatial information is not an

Page 92: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 92

92

absolute piece of contextual information for all problems. Rather, it can be supplementary to

problems that are not inherent to spatial analysis. For instance, crime mapping as illustrated in

the Law Enforcement section of this essay is clearly a location-bound model and therefore spatial

data will not be considered as contextual but is treated as the one of the core independent

variables to predict the dependent variable of crime rate.

The author in (Dey, 2001) proposed a high level abstract architecture for context-aware

application design called the Context Toolkit. It consists of three abstractions of context widget,

context interpreter, and aggregator component to provide features of capture and access of

context, storage and distribution and independent execution for context-aware applications. A

more recent paper published by (Baldauf, Dustdar, & Rosenberg, 2007) described the layered

conceptual framework for context-aware design that resembles the Context Toolkit by (Dey,

2001). The layered conceptual framework is as shown in Figure 25.

Figure 25: Layered conceptual framework for context-aware systems (Baldauf, Dustdar, & Rosenberg, 2007)

Page 93: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 93

93

In the world of internet data, the novel and simple way of tagging is one of the many

ways to provide context to data, which is, a simple way to annotate data with metadata that can

help to classify information.

For instance, tagging on internet articles adds significantly richer and broader information

using keywords such as semantics keywords. The approach of tagging often produces a richer

semantic context to the text than techniques relies only on words extraction from text. This is

because the act of words extraction is done after the fact, which is to say, automated algorithms

are being put in a position to make inferences based on supervised learning technique in order to

make sense out of the text. A common technique in this case is a decision tree which enforces

conformity to a rigid, restrictive and hierarchical categorization format. On the other hand, a

tagging exercise is a human activity, which is to say, the contextual awareness and contextual

information assigned by the person who created the article, and that can be directly translated

into metadata. To that end, metadata information adds a new dimension to data that lack context.

Folksonomy is one way to produce collaborative tags for information classification and

categorization. Tags add contextual information to content data to support a greater level of

human-application communication. The simple means of folksonomy is a great way to allow for

a more accurate information retrieval and response system.

We have established that context is important to information processing within the realm

of information systems. Big Data embodies the ideas and practices of context-aware computing

where the information in silos are not expressive as correlated information in masses. Big Data

extents the notion of what context consists of. Reasonably, for computer systems to predict an

imminent event with high accuracy requires high resolution dataset and context-aware

algorithms. Reasoning and arguing within interpersonal communication requires facts. Ideally,

Page 94: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 94

94

only those facts that are relevant to the discussion at hand. Same is true for computer systems to

reason and argue with data effectively. We are to analyze data that is pertinent and applicable

only to answer the problem in question. Since relevance, pertinence and applicability in

facts/data are subjective measures, Big Data is a means by which we can embody contextual

information. To reasonably derive relevant information associated with analytical directives that

require rich context to be effective, Big Data does provide context for computer systems as the

foundation for context-aware analysis. As such, Data Mining and Predictive Analytics both

operate within the realm of Big Data that can take information processing to the next level. This

is because of Big Data, for its ability to be conceptually closer to the interpersonal form of

communication.

To bridge the gap between Predictive Analytics and Data Mining, the paper by (Kiseleva,

2013) clearly depicted the feedback loop dependency between Predictive Analytics and Data

Mining. In that, the dependency exists in context mining, context modeling and context

integration processes.

Recommendation system is one of the examples in the direct application of context

mining. Many online retailers use recommendation system to make merchandise suggestions to

online shoppers. Context in this case takes many forms and often context is hidden. For example,

an online travel agent company sends marketing emails to potential customers. The company

would target those customers with the aligned purchase intent given the temporal context and

hidden context. The temporal context could be an upcoming public holiday which suggests the

travel availability of the potential customers. The hidden context is the motivation behind the

purchase intent which could be the workplace stress experienced by the customers. Mining both

Page 95: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 95

95

the temporal context and hidden context produce the purchase intent of potential customers

whom might be interested in a leisure trip.

To that end, an architectural description to context mining is depicted in Figure 27 where

the variable set of 𝐶 represents the contextual categories, variable 𝐿 represents a set of individual

learning procedures with the key component of contextual feature set 𝐹, as well as the two

mapping functions of 𝐺 and 𝐻. In short, once the contextual features have been identified, the

function of 𝐶𝑖 = 𝐺(𝐹𝑖) can map features to contextual categories and the function of 𝐿𝑖 = 𝐻(𝐶𝑖)

can map contextual categories to one or more individual learners. This is a high level framework

of what a context-aware application architecture would consist of. This framework was designed

to anticipate the dynamic nature of contextual information.

The design of a context-aware application architecture (Figure 27) shows resemblance of

the Meta-Learning Architecture (Figure 26) described in the paper by (Singh & Rao, 2013) as

well as the context managing framework architecture (Figure 28) by (Baldauf, Dustdar, &

Rosenberg, 2007). The design similarities between the abovementioned architectures further

prove the validity of a fundamental context-aware system design. The context-aware design

architecture also fall under the category of ensemble machine learner or bagging with the

individual learners trained on different datasets with varying levels of techniques and data biases.

Page 96: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 96

96

Figure 26: Meta-Learning Architecture (Singh & Rao, 2013)

Figure 27: An example of context-aware system design (Kiseleva, 2013)

Page 97: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 97

97

Figure 28: Context managing framework architecture (Baldauf, Dustdar, & Rosenberg, 2007)

Context discovery and context integration as described by (Kiseleva, 2013) is necessary

to build context-aware predictive models so as to create context-aware application. While

temporal and spatial information are pertinent to many situations, it is important to clarify that

temporal and spatial based variables are not classified as contextual information. For instance,

time series analysis perceives time as one of the independent variables and crime mapping

perceive location as a key variable to predict crime. However, temporal and spatial data in the

aforementioned examples do not constitute as context element and should be treated as first-class

variables during modeling.

Basic Statistical Methods and Techniques

Many statistical methods and techniques involve searching and measuring the central

tendency of a population as a reference point for other derivative techniques. The central

tendency can take the forms of mean, median and mode to represent arithmetic average, middle

value(s) and highest frequency values, respectively. Most commonly, mean measure is often

used. The central tendency is always required to visualize certain characteristics of a population

at a high level. For instance, the measure of standard deviation (represented in the Greek

Page 98: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 98

98

alphabet sigma Ϭ) in a demographic distribution can display a degree of population skewness in

reference to the Gaussian distribution (i.e. normal distribution), which is based on the Central

Limit Theorem. A Gaussian distribution is a distribution where the mean, median and mode

measures are all equal to one another. It can be visualized as a symmetrical curve called the Bell

Curve as shown in Figure 29 where the population mean (represented in the Greek alphabet µ)

equals to 200.

Figure 29: An example of a normal distribution "Bell Curve"

The measure of standard deviation and variance (Ϭ2) provide a mathematical way to

depict the degree of deviation of a particular sample population using the sample mean (�̅�) from

a normal distribution. To calculate Ϭ2 of a given sample population, the calculation is done based

on the square of the sum of the differences between each data point and the mean, then divide the

result by the number of data points. The mathematical formula is shown in Figure 30.

Ϭ2 =∑ (𝑥𝑖 − �̅�)2𝑛

𝑖=1

𝑛

Figure 30: Variance formula

0

50

100

150

200

250

300

350

Children Preteen Teen Adult Middle Age Senior

Number of People Per Age Group

Page 99: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 99

99

Once Ϭ2 is calculated, to calculate Ϭ is simply the square root of Ϭ2 which are shown in

Figure 31 and Figure 32.

Ϭ = √∑ (𝑥𝑖 − �̅�)2𝑛

𝑖=1

𝑛

Figure 31: Standard Deviation formula 1

Ϭ = √Ϭ2

Figure 32: Standard Deviation formula 2

The example shown in Figure 29 is an example of univariate analysis where there is only

one independent variable called “age group”. On the other hand, multivariate analysis allows the

use of two or more independent variables. A bivariate analysis deals with analysis of two

variables and measuring the correlation in bivariate analysis is essentially the cornerstone of all

Predictive Analytics methods.

Linear correlation coefficient (represented in 𝑟) is a mathematical way to discover hidden

relationship and detect any correlation between two variables. 𝑟 is calculated by the sum of each

product of the differences between two variables and the respective means, divided by the

product of the standard deviation of the two variables and multiple the result by the number of

data points. The formula is shown in Figure 33.

𝑟 =∑ (𝑥𝑖 − �̅�)2(𝑦𝑖 − �̅�)2𝑛

𝑖=1

𝑛Ϭ𝑥Ϭ𝑦

Figure 33: Correlation Coefficient formula

The value of 𝑟 exists between the range of -1 to 1 where 0 represents absolute non-

correlation, positive number refers to positively correlated and negative number refers to

negatively correlated. The direction (i.e. positive or negative) of the correlation refers to the

directional movement of the two variables. If variable 𝑥 increases while 𝑦 decreases, there is a

Page 100: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 100

100

negative correlation exists between 𝑥 and 𝑦. Similarly, if both 𝑥 and 𝑦 increases or decreases

create a corresponding change against each other, then it is said that there is a positive

correlation between 𝑥 and 𝑦. Therefore, the result of −1 < 𝑟 < 0 represents the degree of

negative correlation and 0 < 𝑟 < 1 represents the degree of positive correlation of the two

measured variables.

There are many versions of the 𝑟 mathematical formula but the underlying promise is the

same. It is used to compute and output a representative numerical value that corresponds to the

level of correlation between two variables. Many statistical modeling techniques rely on this

concept. A simple linear regression modeling based on least squared method is one such

example.

All of the predictive statistics used in statistics is inferential statistics for modeling data

that are high in randomness and uncertainty, a measure that accounts for entropy which is

inherent to the data being modeled. A common way in scientific and academic research for

statistical hypothesis test is to confirm hypothesis based on experimental analysis involving

statistical methods. Covariance, correlation and the measure of standard deviation are methods

that generate information apposite to be inferred to a hypothesis or established theory. All models

are subjected to model validation to ensure high predictability and inference-making ability. A

general approach to model validation is to train model with a randomized subset of the sample

data and use the remainder sample data to validate. Splitting sample data for both purposes in

model training and validation allows researchers to maximize the value of the available data.

This dual-purpose practice can take the forms of holdout and subsampling, cross-validation and

bootstrap methods.

Page 101: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 101

101

Data Mining Methods and Techniques in Predictive Analytics

Under the Predictive Analytics section of CHAPTER II, we briefly discussed the various

methodologies in Data Mining and where do Predictive Analytics fit in within the Data Mining

archetype. The strong focus on data visualization methods in Predictive Analytics is evidently

reflected on the popularity of scatterplot charts across many industries. The commonplace

scatterplot chart is a visualization tool for linear and non-linear regression-type analysis.

Classification is also another highly regarded methodology in Predictive Analytics for its ability

to distinguish data classes by labelling them for future inferences. This process is often called

concept classification. Both regression analysis and classification analysis are prevalent

methodologies in machine learning. Many techniques derived from them such as neural network

based methods (Mathewos, Carvalho, & Ham, 2011), clustering methods (Zeng & Huang, 2011)

and dimensionality reduction methods (Vlachos, Domeniconi, Gunopulos, Kollios, & Koudas,

2002). For example, the simple decision tree model consists of both classification and regression

variants called classification tree and regression tree, respectively. Regardless of the chosen

methodology and method, the historical context in Data Mining is crucial to the development of

any predictive model. Mining historical information of known facts allows a method to find

similar and probable outcome for the unknown. This concept underpins the fields of Data Mining

and Predictive Analytics as well as all of their derived methodologies and methods.

Classification

To understand how useful classification methodology is in Predictive Analytics, one must

presume inference is a requisite of prediction and therefore any technique that is grounded in

logical inference can be used for Predictive Analytics. Classification at its core is an inference

engine because it follows a logical order of operations based on learnt rules to segment data into

Page 102: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 102

102

common constituents. Classification models that use training data to generate model rules are a

type of supervised learning. Most classification techniques are supervised learning based and can

be modelled based on techniques such as decision tree classification, naïve Bayesian

classification, Bayesian belief networks classification, support vector machines (e.g. kernel

approximation) classification, nearest neighbors classification, stochastic gradient descent (SGD)

classification, multiclass and multilabel classification.

Many of the aforementioned classification techniques are also used in solving regression

problems, they are models that can be applied to both classification and regression problems. The

major difference between classification and regression is that classification deals with discrete

and categorical dataset while regression deals with continuous and numerical dataset.

Bayesian belief networks is a classification technique that can be used to recover

information about the structure of statistical dependencies among variables, quantifying

probabilistic conditions such as 𝑃(𝑌|𝑋). The probabilistic conditions can be represented in a

graph network of nodes, marking the connections between events and belief variables that are

either conditionally dependent or structurally independent. The result can be observed and

quantified as shown in Figure 34 and illustrated below:

Given 𝑋 → 𝑍 → 𝑌, variable 𝑌 is conditionally dependent of 𝑋 through 𝑍 such that

𝑃(𝑌|𝑍, 𝑋).

Given 𝑋 → 𝑌, 𝑋 → 𝑍, variable Y is conditionally dependent only on X and

structurally independent of Z such that 𝑃(𝑌|𝑋).

In Bayesian belief network, the probabilistic based conditional rules are captured within a

conditional probability table (CPT) for each variable that presents a given 𝑃(𝑌|𝑋) condition.

Given that, a conjoin condition can be calculated mathematically through the multiplication of

Page 103: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 103

103

𝑃(𝑌|𝑋) conditions to arrive at a joint probability distribution under the following Bayesian

inference formula, for the condition 𝑋 → 𝑌: 𝑃(𝑌|𝑋) =𝑃(𝑋|𝑌)𝑃(𝑌)

𝑃(𝑋) where 𝑃(𝑋) is the product of

the conditional dependence variable set (including the parents of 𝑃(𝑋)) that 𝑌 dependents on.

Figure 34: A Bayes net for the medical diagnosis example (Patterns of Inference, 2014)

Support Vector Machine (SVM) is also a frequently used classification technique. SVM

deals with both the linear and non-linear classification problems. SVM concerns with delineating

data points with the use of hyperplanes while maximizing the margins (i.e. the distance between

the two hyperplanes). To that end, SVM shares the same principle of clustering method which

will be discussed in the Clustering section.

Regression

Regression analysis shares many predictive properties of classification which is ideal for

Predictive Analytics for its ability to normalize variables through function approximation.

Page 104: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 104

104

Visually speaking, regression is a way to describe the dots-and-lines spatial relationships where

the dots represent the data points and the lines represent the relationships amongst the data

points. An example of visualized linear regression method is shown in Figure 35. In short,

regression analysis allows linear and non-linear fitting of scattered independent variables through

mathematical techniques. The result of regression analysis is often visualized via scatterplot

chart. Regression methods such as non-linear regression, logistic regression, support vector

machines regression, stochastic gradient descent regression, Gaussian processes regression,

decision tree regression and isotonic regression can provide forecasting capability visually and

mathematically.

Figure 35: A visualized example of linear regression (Natural Resources Canada, 2012)

In the world of Predictive Analytics, logistic regression is a form of regression analysis

that is a highly regarded method for prediction due to the probability based sigmoid function of

Page 105: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 105

105

𝑃 =1

1+e−β𝑘 where 𝑃 is the calculated probability and the β is the matrix vector of 𝛽𝑘 = 𝑎 +

𝑏1𝑥1 + 𝑏2𝑥2 + 𝑏3𝑥3 + ⋯ + 𝑏𝑘𝑥𝑘 + 𝑒 which comprises a linear equation and error (𝑒) parts. The

𝑃 produces a logistic “S” shape curve visually between the value of 0 and 1. Rather than

producing a continuous numerical dependent variable as linear regression does with ordinary

least squares method, logistic regression produces a binary response that is best to describe any

dichotomous phenomenon and thus it is suitable for classification problems.

Clustering

In most cases, the classification and regression methodologies involve supervised

learning methods. Supervised learning refers to model building with pre-labeled training dataset

where manual intervention of data classification was performed on the training dataset, which is

to say, supervised learning models learn from examples prepared by human.

Clustering is an unsupervised learning methodology without the need for human

intervention for training dataset, which means that, clustering methods operate directly on live

data. The unsupervised learning nature of clustering is both advantageous and limiting.

Unsupervised learning techniques generally produce inferior predictive power than supervised

learning techniques. However, unsupervised learning such as clustering benefits from

autonomous operation, relying entirely in intrinsic data properties such as central tendency and

density to filter and assemble data based on the relative relationships (e.g. Euclidean distance).

Very often, clustering serves as a great first step for data analysis. Clustering methods such as k-

Means clustering, Gaussian mixtures, hierarchical clustering, spectral clustering, mean-shift

clustering, DBSCAN clustering and affinity propagation clustering, all depend on the pre-

existing relationships between data points.

Page 106: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 106

106

K-Means clustering is the one of the most common approach in clustering. K-means

clustering algorithm expects an input of 𝑘 where 𝑘 is the expected number of cluster groups

which the algorithm would produce. A group of data sharing the nearest mean value is

algorithmically determined to form a single cluster relative to the expected number of clusters.

The k-means clustering method would continue to form clusters and rebalance previously formed

clusters until 𝑘 number of cluster groups have been created. This iterative process is shown in

Figure 36. This clustering technique has strong dependence on the statistical techniques

discussed in the Basic Statistical Methods and Techniques section.

Figure 36: Clustering of a set of objects using the k-means method; for (b) update cluster centers and reassign objects

accordingly (the mean of each cluster is marked by a C) (Han, Kamber, & Pei, Data Mining: Concepts and Techniques, Third

Edition, 2011)

Artificial Neural Network

Artificial neural network mirrors the biological design of a human brain in which neuron

is emulated as perceptron containing a function called activation function or transfer function.

The functions are encapsulated within a multi-layer topology of interconnected perceptrons

consisting of input layer, hidden layer and output layer. This simple definition allows for

artificial neural network to spread across the entire range of machine learning discipline.

Artificial neural network can be used in areas that require non-linear prediction,

prediction for phenomenon with covariate relationships, data classification, feature extraction

Page 107: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 107

107

(e.g. PCA and Factor analysis), data compression and general application in image processing

(e.g. de-noising and recognition). Consider the most basic form of artificial neural network in

non-linear machine learning called the backpropagation neural network. The backpropagation

neural network is based on the multilayer feed-forward neural network topology (Figure 37) with

the supplemental ability to back propagate across sub-layers within the hidden layer, as a means

to update previously learnt weights and bias. Backpropagation neural network is a supervised

learning method which means that training dataset with targeted labels are provided to train the

perceptrons’ connection (i.e. edge) weights and unit (i.e. node) bias. Since the neural network

topology allows for many-to-many connections between perceptrons, it structurally defines each

perceptron from one layer passing output value to the perceptron in the next layer as an input

value in order to form a complete network. Each perceptron makes a decision based on output

from the previous perceptron and adds their connection weight value and bias value to derive a

binary response of 1 or 0. A collection of these perceptrons forms a network of nodes that

produce a collection of binary output values similar to the way human brain adapts and learns

new knowledge based on synapses that occurs between axon and cortical neuron.

Page 108: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 108

108

Figure 37: Multi-layer feed forward neural network (Han, Kamber, & Pei, Data Mining: Concepts and Techniques, Third

Edition, 2011)

One way to implement linear perceptron is to use binary threshold neuron formula as

shown in Figure 38 to produce a binary response (i.e. 0 or 1, true or false) for all outputs of other

input neurons based on weighted sum, where 𝑤𝑖 is the weight per neuron connection, 𝑥𝑖 is the

input value, 𝑏 is the bias value and 𝑦 is the neuron output based on the calculated weighted sum

of 𝑧. The value of 𝑦 can be used to determine if an output will be directed to the next perceptron

for a given calculation. For instance, if 𝑦 = 1 then perceptron will send output to the next

perceptron; if 𝑦 = 0 then perceptron halts output. A collection of these small decisions thread

together to form a neural network with zones (i.e. areas of selected perceptrons). The zones

within a neural network represents pathways based on learnt data which would directly influence

Page 109: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 109

109

future outputs based on similar inputs (i.e. training dataset). This representation and mechanism

work much like the synapses of a human brain. The synapses become stronger overtime when a

particular part of memory has been exercised more. Other implementations include sigmoid

neuron and stochastic binary neuron and both implement logistic regression for binary output.

𝑧 = 𝑏 + ∑ 𝑥𝑖𝑤𝑖

𝑒

𝑖=𝑠

𝑎𝑛𝑑 𝑦 = {1 𝑖𝑓 𝑧 ≥ 00 𝑖𝑓 𝑧 < 0

Figure 38: Binary Threshold Neuron formula

Backpropagation artificial neural network is an inverse function of a typical feed-forward

artificial neural network model for supervised learning purpose. To train a backpropagation

model, the error derivative is derived from the delta of the predicted value and the actual (i.e.

labelled) value which is used to update the model. This process occurs continuously and

iteratively.

Researchers have made many improvements to artificial neural network in hope of

enhancing the model to better solve real world problems. One such method is to hybridize with

fuzzy logic to create neural fuzzy network system (Nirkhi, 2010). Contrary to the standard

sigmoid function that produces only binary response, the neural fuzzy network system produces

a degree of measure based on fuzzy rules and fuzzy sets.

Conclusion

In concluding this section, we have discussed the basic premise of Predictive Analytics in

which Predictive Analytics, as part of Data Mining, is realized by the algorithms that implement

the machine learning models. The machine learning models are underpinned by the century old

disciplines of statistics and mathematics. The importance of statistical analysis and the role it

plays in machine learning are also discussed.

Page 110: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 110

110

The learn-to-predict paradigm of Data Mining dominated most of practical applications

in the form of supervised learning techniques and methods, the algorithms that are derived from

these machine learning methods has the capacity to infer based on data. Inferences made by these

algorithms are logical and evidence based and the result is justified empirically given the

deduced, features extracted, dimensionally reduced and correlated dataset. This lead to the

development of the overarching methodologies of classification and regression which provide

data labeling and concepts grouping.

Conceptually, regression methods such as the simplest form of regression, ordinary least

regression, is identical to the classification methods such as decision tree, in which, the data are

normalized and categorically divided by a single label (e.g. classification) or a line (e.g.

regression). The label or line abstracts away the complex relationships between the independent

variables as well as the conditional expressions and mathematical calculations that best describe

the data being analyzed.

It is when we transpose the inference into prediction based on learnt data, that we make

the leap from descriptive and explanatory analysis to prescriptive forms of analysis. Another way

to look at Predictive Analytics through the lens of statistics is the transformation of independent

variable into predictor variable. Finding the most fitting independent variables to infer to a

dependent variable, and through which to derive a predicted outcome. This is the reason that the

modeling techniques are similar between retrospective and prospective ends of the analytics

spectrum.

The consensus amongst researchers at the current state of Predictive Analytics is that the

differentiator in predictive performance came from the quantity and quality of the data during

supervised learning rather than favoring the algorithms and models themselves. This is where

Page 111: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 111

111

Big Data becomes a significant interest in future research. The pinnacle of predictive modeling

presently arrived at the Ensemble modeling methods which will be discussed in the next section.

We conclude this section with the understanding that Predictive Analytics problem is a

statistical problem and that statistical problem can be solved by many of the aforementioned

techniques.

Page 112: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 112

112

CHAPTER V

ISSUES, CHALLENGES, AND TRENDS

Introduction

It is evident that the techniques and methods described in CHAPTER IV are an integral

part of Predictive Analytics. The design of models and advanced techniques are important to

improving predictive accuracy. However, predictive modeling will inevitably be optimized to

reach a pinnacle where the models and methods themselves will no longer provide a measurable

lift against competing models, resulting in a predictive performance plateau. In such cases, no

gain in predictive power can be achieved simply by improving the model itself or by applying

other advanced techniques. The tipping point for better predictive performance tends to give way

to the data that feed the models. In which case, the characteristics of Big Data provide an

abundant source of training data for predictive modeling.

The quality of the data certainly plays a critical role in the predictive power of a given

model. Data quality is multidimensional where data recency, data volume, data variety and data

veracity are the basis to provide a timely, abundant, unbiased and accurate training dataset for

any given model. If the training dataset is of poor quality, no predictive model can excel and the

result could even be inferior then a random guess. If the data are bias, in a sense that they

represents only a minority of a subset of the common view, no predictive model could predict

any result beyond the boundaries set by the data that inherently limited the model. This is the

classic “garbage in garbage out” saying or the “the output is only as good as the input” principle

of any computing systems. Human and computing systems are bound by the same principles,

predicting acceptably without bias and flaws equate to learning without knowing, a paradoxical

Page 113: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 113

113

conundrum that no human or any computing system can solve because we can only process

within certain confines that are restrained by the presently available and accessible information.

Uncommon events are rare events that are difficult to predict. Through the discussion

thus far, we underscored the difficulty to predict without inference and to infer without evidence.

The learning models reinforce our collective reliance on data as a single means to provide

evidence for inference that leads to prediction. Committing a type I error (i.e. false positive) on

predictions might not be an issue if the error results in only minor inconvenience. For instance,

incorrectly forecasted meteorological conditions such as rainfall in an urban setting might not be

impactful to the lives of the individuals living in the area. However, generating a false negative

on rare but catastrophic event, thereby committing a type II error, could have disastrous

consequences such as major natural disaster (e.g. earthquake), terrorist attack and major

economic downfall. Balancing between a tolerable level of false positives and discovering

momentous false negatives marks the major obstacle that statistician and mathematician have

been battling with since.

Performing statistical modelling will inevitably lead to the conundrum of goodness of fit.

An important measure used by researchers to assess if the developed model is describing the

observations (i.e. data) between the dependent variable and the independent variable adequately.

This is done by measuring what we expect to see (i.e. hypothesis) versus what we actually saw

(i.e. observations). Some of the methods for goodness of fit measure include Pearson's chi-

squared test and p-value method.

When a model is unable to sufficiently describe the data, it is said the model is

underfitting the data. For instance, using regression analysis as an example, describing a

dependent variable with a binary response using linear regression is undesirable and would

Page 114: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 114

114

grossly underfit the data due to high residuals (Lesson 19: Logistic Regression, 2014). Using

Figure 39 as an example, suppose there is a significant income disparity between female and

male where the income of male is higher than female. This would result in a binary response

dependent variable (i.e. gender) for any given income. A linear regression in this case is not a

descriptive method and would result in a straight line penetrating two clusters of data points as

shown in Figure 39. A good model in this case would be the logistic regression model as

discussed in the Regression section.

Figure 39: An example of binary response dependent variable (x=income, y=gender)

Page 115: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 115

115

On the contradictory, in the predictive realm of statistics where a model can overfit the

data rather than underfit which can cause diminishing predictive power of a model for unseen

events. Reasonably, fitting the model fully and perfectly is great for explanatory and descriptive

work. If researchers can impeccably identify the variables for a given outcome, fitting the model

exactly to the data is desirable. However, the opposite is true for predictive modeling. Predictive

model should not fully and perfectly fit the data. This is because an overfitted model is a model

that lacks agility and therefore possesses minimal predictive ability. In this case, the margin of

error in predictive model represents the degree of flexibility in dealing with unseen events.

To that end, a functional predictive model should strike a balance between overfitting and

underfitting of the training data because certain level of rigorous and nimbleness must be present

for prediction. While the majority of the model designs are to make parallel to the training data,

it is also important to consider outliers and odd cases, a built-in lenience for exceptions is needed

to build capacity for prediction.

Big Data Issues, Challenges and Trends in Predictive Analytics

Using the Google Flu Trends example illustrated in the Preliminary Discussion section of

CHAPTER IV. The example illustrated the achievement of a higher level of predictability when

Big Data was introduced into the equation. However, a recent study by (Lazer, Kennedy, King, &

Vespignani, 2014) suggested the underlying Google Flu Trends prediction mechanism led to an

overestimation of the prevalence of flu between year 2011 and 2013. The study is a reminder for

us of how Big Data might not actually deliver what it has promised.

Realizing the association between prediction and statistics and deriving predictions

means solving classification problem, the Big Data paradigm shift within the domain of

Predictive Analytics becomes an apparent and significant factor. This is because, the Big Data

Page 116: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 116

116

characteristics provide an abundant sources of already classified data that are available and

accessible on the internet. Supervised machine learning with Big Data reaps the benefits from the

attributes and characteristics of Big Data, that is, the volume, velocity and variety aspects. The

volume aspect of Big Data represents size, scale, dimension and the amount of data collected

which is an important enabler for prediction. Prediction performs intrinsically better with more

data, which is to say, the high data volume corresponds to more granular class labels for

classification or high data points for regression function approximation.

In the broadest sense, more data is always better for prediction since data adds context to

the inductive nature of all supervised training methods. Furthermore, more data adds extra

dimensionality to the set of independent variables which allows for high dimensional data

analysis to account for multi-factors and latent variables. While additional independent variables

do not naturally become high value predictor variables, the probability of discovering predictor

variables increases as more independent variables are available through Big Data.

Machine learning methods such as backpropagation artificial neural network requires vast

amount of training dataset to be effective due to the recursive process of the update of weight

values throughout the network topology. The argument for how the high volume aspect of Big

Data can benefit Predictive Analytics is identical to that of scientific experiment. Scientific

experiment requires sufficient number of observations in sample data to be representable in a

population. Having insufficient observations likely decrease external validity which impair

inference-made outside of the sample, that is to say, the model is unable to sufficiently make

prediction beyond its sample data. Small data makes the case for biased data; big data increases

the chance for an evenly distributed randomized sample and reduces the risks for data bias. Big

Data can avoid the overrepresentation and underrepresentation of any given dataset due to the

Page 117: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 117

117

fact that, the sample size is approaching the population size. Therefore, Predictive Analytics

within the context of Big Data, at the very least, can benefits from the unprecedented high

volume of training data. This is because, more is not just more, more is different (Anderson,

1972).

Consider the high velocity aspect of Big Data, prediction demands real-time or near real-

time data to account for rapid shift in data context. Prediction model becomes uninformed when

operating on antiquated data that lacks data recency. The future will always remain a degree of

uncertainty notwithstanding the advanced analytics techniques we have discussed this far.

However, uncaptured events and uncollected data are missed opportunities. The real-time or near

real-time data feeds enable the prediction model to sift through data into the future relative to

model that relies on hours or days old data. The benefits are two-folds. Firstly, model operating

on timely dataset avoids making prediction on events previously occurred, thus maximizing

computing resources on the most relevant context. Secondly, achieving real-time data processing

means processing against the current rather than the past, an important distinction for

applications such as algorithmic trading.

Finally, the variety aspect of Big Data fuels the arguments made on the volume aspect of

Big Data in which wide-ranging data types and data sources add data volume to the model

training data. Also, the data type and data source variety in Big Data, in and of itself, are

metadata that can be contextualized for Predictive Analytics. This is because significant

differences exist between different data types as well as the data sources which are contextual

self-descriptive, to a certain degree. For instance, consider the law of evidence, significant

differences exist between a witness written deposition and a video evidence that captured an

event of a crime. Admissible video evidence would prove to be a powerful piece of evident in a

Page 118: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 118

118

court case. This is because video evidence generally provides a richer context that is rarely

available in written deposition based on malleable human memory. In that sense, the metadata

for document and video data type, is the fact that video would carry a higher degree of evidential

weight value than a written deposition in the case above. Furthermore, news information that

came from a personal website generally carries less weight of trustworthiness than a reputable

news outlet such as the Canadian Broadcasting Corporation (CBC). Therefore, the variety aspect

of Big Data not only provides a diverse data types and sources for data, but also provides the

needed context that is often overlooked during predictive modeling.

Trend in Big Data Application

Processing textural data and visual data require vastly different techniques to perform.

Processing a written document would require text analytics (Kimbrough, Chou, Chen, & Lin,

2013) such as natural language processing techniques that involves deep linguistic processing

methods. However, processing images of objects in photo requires object recognition techniques

that involve edge detection methods while processing video data requires motion detection

techniques. All of which are some forms of features extraction techniques to address very

different problem domains that very often involve data dimension reduction techniques for

information processing. The diversity in data types unquestionably created challenges for

researchers, on one hand, more is different; on the other hand, more is complex and difficult.

The assorted data processing techniques mentioned in previous chapter existed prior to

the advent of the Big Data era. However, the existence of Big Data accelerated the growth in the

application of the aforementioned techniques which are heavily used in predictive modeling. As

in the study by (Jarrold, et al., 2010), using data mining, machine learning and text analytics

techniques to predict future patient brain health in the areas of: cognitive impairment, depression

Page 119: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 119

119

and pre-symptomatic Alzheimer’s disease. The authors of the study applied multiple techniques

on patient language expressions, one of which was supported by the Linguistic Inquiry Word

Count (LIWC) software (James W. Pennebaker; Roger J. Booth; Martha E. Francis, 2014) by

performing text analytics on transcribed audio interview data. The authors concluded the study

with the suggestion to further exploit web intelligence (i.e. Big Data) to further research the

application of text analytics in the disease prevention domain. The potential was highlighted by

the success prediction rate between 73% and 97% in the study.

In another study by (Schwegmann, Matzner, & Janiesch, 2013), the authors facilitated

predictive event-driven process analytics under the Complex Event Processing (CEP) technology

by exploiting event log data to improve business functions, an architectural approach to integrate

Predictive Analytics with CEP. CEP is a broad term where event data is described as “anything

that happens, or is contemplated as happening” (Llaves & Maué, 2011). CEP thus refers to the

processing aspect of the ubiquitous event data. The nature of event data ties directly to the

discussion of Big Data.

Log event data are often overlooked and underutilized, however, data embedded in

machine generated log files are often context rich that can help other aspects of a given

operation. To combine Predictive Analytics with CEP is to use Predictive Analytics with Big

Data to improve processes, especially the velocity aspect of Big Data.

Page 120: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 120

120

Figure 40: Value of knowledge about event (Fülöp, et al., 2012)

As shown in Figure 40, the value of an event depreciates over time on the application of

Predictive Analytics. Ideally, proactive actions can be administered on an impending event.

Beyond that point, it would become a reactive action that is still high in practical value but it is

limited within a sub-second period; a period of time immediately follows an event. To that end,

the authors in (Fülöp, et al., 2012) proposed a conceptual framework for combining CEP and

Predictive Analytics as depicted in Figure 41. In that, the framework consists of predictive event

processing agents within a predictive event processing network to synergize both aspects of

Predictive Analytics and real-time processing of event data.

Page 121: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 121

121

Figure 41: CEP-PA conceptual framework

Beyond the realm of complex event processing, the multi-purpose application of

Predictive Analytics was also researched within the context aviation surveillance Big Data.

Specifically, the Aircraft Situation Display to Industry (ASDI) data generated over the years

ranging from flight arrival information to oceanic report, from plane maintenance history to

flight route information. The study was discussed in (Ayhan, Pesce, Comitz, Gerberick, &

Bliesner, 2012). The authors applied Predictive Analytics in the aforementioned massive

surveillance Big Data accumulating at rate of approximately 1GB per day. The collected

information combined with the developed system illustrated by the authors, can answer practical

concerns such as flight route optimization. The author concluded the paper with plan to include

meteorological conditions information to further enhance the underlying model supported by

SPSS Modeler software.

Page 122: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 122

122

Predictive Analytics Issues, Challenges and Trends

A study by (Niculescu-Mizil & Caruana, 2005), the authors quantitatively identified 10

common supervised machine learning methods, that are, boosted tree, support vector machine,

boosted stumps, naïve bayes, artificial neural network, bagged tree, random forests, logistic

regression, decision tree and memory-based learning. Each method possesses a degree of bias

within its structured definition of the methods, resulting in a distortion in prediction. The authors

proposed two methods (i.e. platt calibration and isotonic regression) for supervised machine

learning calibration in an attempt to reduce the distortion produced by the learning methods.

The result suggested that the calibration methods do in fact produced a measurable

improvement in the learning methods (e.g. SVM, boosted stumps) while bagged trees and

artificial neural network did not produce any observable improvement. The best learning

methods in terms of minimizing errors are: boosted tree with platt scaling calibration, random

forest with platt scaling calibration, non-calibrated bagged tree and non-calibrated artificial

neural network. The study benchmarked the 10 most common methods and the result do suggest

a plateau in supervised learning model performance, particularly, both bagged tree and artificial

neural network performed worse after calibration. Although not mentioned in the study, a few of

these models fall under the category of ensemble methods which further divided into two

subcategories: bootstrap aggregating (e.g. random forests) and boosting (e.g. boosted tree)

methods.

The Ensemble Approach

To appreciate the ensemble approach is to embrace diversity in aggregate where

collective intelligence can be achieved. Ensemble based supervised machine learning methods

are gaining acceptance and exemplified by the one million dollars Netflix Prize competition. The

Page 123: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 123

123

goal of the event was to predict movie rating of Netflix users, a form of collaborative filtering

problem. The top performing team in the event had all employed ensemble based models (Siegel,

2013). An interesting series of events occurred during the competition where competing teams

began forming alliances amongst each other to compete with other rival alliance teams. Drawing

on the expertise of the models created by other teams to create an ensemble model with meta-

learning capability which allows automated model selection based on the strengths and

weaknesses existed within the individual models.

The event concluded with BellKor’s Pragmatic Chaos ensemble team as the winner of the

Netflix Prize event (Koren, 2009) which was an alliance team formed amongst BellKor,

BigChaos and PragmaticTheory teams. The ensemble model developed by the team produced a

10% improvement in movie rating prediction over the Netflix’s internal established method. The

winning model employs a Gradient Boosted Decision Trees (GBDT) method as discussed in the

related literature by (Ye, Chow, Chen, & Zheng, 2009). GBDT is a boosting based ensemble

decision tree method consisting of multiple decision trees.

Ensemble methods are gaining momentum, it is because the underpinning meta-learning

idea of combining weak leaners to make a strong leaner means existing model performance can

be improved by aggregation. For instance, using AdaBoost (Freund & Schapire, 1999) as an

example, the algorithm maintains a set of weight values over a set of base learners and the

weight values were adjusted over time through supervised learning. The AdaBoost operation not

only smooth out the biases exhibited within the constituent leaners, but also matches the model

with any given domain problem (i.e. predictor) that has the best empirical evident (i.e. in terms

of previous predictions) as the best performer. Therefore, boosting based method such as

AdaBoost, can learn from the result of the base leaners to play to their strengths in an assembly

Page 124: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 124

124

of leaners. The strength of ensemble based method is evidently established by the empirical

evident in (Ye, Chow, Chen, & Zheng, 2009) and in (Siegel, 2013).

The Concept Drift

Another advantage to boosting method is the dynamic nature of the technique itself

which help to alleviate the challenges posed by concept drift (Venkatesan, Krishnan, &

Panchanathan, 2010). Concept drift defines the phenomena involving changes in the predictive

nature of the independent variables used in the underlying data and model. Concept drift

describes data that exhibits a shift in variable relationship in concept, which is different from the

training data used during the model’s supervised learning process. In other words, the training

data used to train a model no longer representing the current data being processed by the model.

This is understandable, since the training data is always a subset of the population data

and therefore poses a chance for concept drift. As well, the population data is dynamic in nature

and not static. Consequently, a model that was trained with sample data must relearn at some

point to be in alignment with the new sample data. Many techniques have been developed by

researchers to detect concept drift as discussed in (Masud, Gao, Khan, Han, & Thuraisingham,

2009). AdaBoost in this case, to a certain degree, possesses inherited properties to guard against

concept drift due to the dynamic weight value which represents past performance of a given

learner. In this case, if a base leaner exhibits a sudden drop in predictive performance relative to

other base leaners, a concept drift might have occurred and thus requires model rebalancing and

possible model retrain.

Trends and Advancements

Beyond the consortium, pooled or ensemble methods, other advances in academia and

scientific communities are thriving in the machine learning domain. Particularly, the Python

Page 125: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 125

125

based open source communities (e.g. SciPy and Scikit). A list of Python based libraries were

identified in the Python Related Statistical Libraries section under APPEDIX B. Notably, the

scikit-learn python machine learning library which offers a comprehensive collection of ready-

to-use machine learning algorithms as identified in the Data Mining Methods and Techniques in

Predictive Analytics section of CHAPTER II.

On the image processing side, scikit-image python image processing library was created

to support specific object and facial recognition applications. Other non-Python based library

includes the java based Apache Mahout (What is Apache Mahout?, 2014) and the C++ based

Dlib library (Machine Learning, 2014). Also, the R platform (The R Project for Statistical

Computing, 2014) and the STATISTICA products (StatSoft, 2014) by StatSoft have a long

history in academia and commercial space. They are used by many researchers and data

scientists for statistical modeling and analysis work. Other open source software supporting

Predictive Analytics application was identified in the Open Source Predictive Analytics and Data

Mining Tools section in APPEDIX B.

On the Big Data side, Apache Hadoop is the most recognizable and dominant opens

source solution to the Big Data problem as discussed in The Apache Hadoop Platform section of

CHAPTER II. This leads to many commercial adaptations of Apache Hadoop platform, some of

the notable Hadoop distributions are: Cloudera, Cloudspace, EMC Greenplum, Hortonworks,

IBM BigInsights Enterprise Edition and Think Big Analytics. In the non-Hadoop proprietary

commercial space, includes Oracle Database (Oracle Inc., 2013) and a host of companies listed

in (Gartner, 2014) for Business Intelligence and in (Gartner, 2013) for DBMS.

On the infrastructure side of discussion, traditional DBMS was designed with

transactional processing in mind and does not scale well to support the modern enterprise data

Page 126: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 126

126

warehouse, as discussed in The NoSQL Solution section of CHAPTER II. Executing these

functions in the database, closer to the data, can avoid data movement and thus accelerate

analytics performance and increase throughput. Further, centralizing analytics in the database

facilitates version control, reduces duplication, and extends the DBMS’s management, security,

and auditing infrastructure to analytical data and functions.

The direction to in-database analytics architecture was realized by the discussion on EMC

Greenplum (EMC Inc., 2012) study under the PMML Enabled Architecture section of

CHAPTER II. Other competing commercial products in the in-database analytics space are IBM

Netezza (IBM Inc., 2012), SAS (SAS Inc., 2007), Teradata Database (Grimes, 2012) and Oracle

Database (Oracle Inc., 2011). As previously mentioned, the R platform has a long history and it

is pervasive amongst practitioners, which attracted many commercial and open source products

to integrate their products with the R platform such as Teradata Database (Teradata , 2013) for

in-database analytics using R.

IBM Watson famously outdid top human contestants in the game of Jeopardy! (Siegel,

2013). The event serves as a practical example of Predictive Analytics with Big Data in action.

IBM Watson utilized text analytics and ensemble models along with internet data to analyze

question (in a form of answer in Jeopardy!) phrased in human language, to predict the most

likely answer by ranking the top candidate answers based on their probability scores. Of course,

the Jeopardy! challenge was meant to be a showcase of IBM Watson’s ability to process massive

information. One of the first commercial applications of IBM Watson was to support healthcare

preapproval decisions based on clinical and patient data (IBM Inc., 2013), applying the same

techniques and methods used on Jeopardy! on practical application of management decision

support.

Page 127: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 127

127

Another notable trend in predictive modeling is the Uplift Model (Siegel, 2013). Uplift

model has the practical properties of finding the differentials in class/label when an intervention

is introduced, which is to say, solving the problem of △ 𝑃(𝑌|𝑋) = 𝑃(𝑌𝑖|𝑋𝑖) − 𝑃(𝑌𝑖+1|𝑋𝑖+1)

where △ 𝑃(𝑌|𝑋) is the probability of delta of the effect between the original class/label and the

intervention. Also, to measure a combined effect of a treatment, we can perform 𝑃(𝑌|𝑋) =

𝑃(𝑌𝑖|𝑋𝑖) +△ 𝑃(𝑌|𝑋).

For instance, a popular example amongst uplift modeling business application was that,

marketing professionals often interest in knowing the persuadable individuals to concentrate

targeted marketing efforts then to spend resources on the unlikely individuals who are

unpersuadable or less-than-effective-persuadable. In that case, discovering the persuadable

individuals is also a specialized classification problem because focuses would be given to those

who are classified as persuadable based on commonly shared predictors. Since uplift modeling is

dealing with classification problem that can be solved by classification methodology. Association

rules such as basket analysis is certainly a logical choice given the problem domain, however,

decision tree method is one of the simplest form of classification techniques that is simple to

understand and implement in most cases.

In (Rzepakowski & Jaroszewicz, 2011), clinical trials often involve control group in

order to observe the isolated effect of a treatment. Uplift modeling is also applicable in this case.

Uplift modeling not only helps to predict the differences between treatment and non-treatment

groups, but also allows researchers to predict the lift from single or multiple treatments applied to

treatment group(s). Furthermore, uplifting modeling can model the effects on each action (i.e.

treatment) applied as well as measuring the lift in the degree of effectiveness per treatment or per

a series of treatments. This is an example of how uplift modeling can be incorporated into

Page 128: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 128

128

clinical trials to better support and to more accurately predict clinical trial successes (Jaskowski

& Jaroszewicz, 2012).

Ethical Concerns and Issues

The ethical concerns overlap both the problem domains of Big Data and Predictive

Analytics. In this context, the concern for Big Data stems from the data collection and storage

aspect of the problem domain while Predictive Analytics matters in the implementation problem

domain. That is, we must look at the Big Data side of our concern prior to investigating in

Predictive Analytics. Big Data, by definition, increases the data resolution and thus contains

details that are of concerns to some.

Privacy is a basic human right enforced by laws in most industrialized countries

including Canada. Canada established the Privacy Act (Government of Canada, 1985) to

complement the Access to Information Act (Government of Canada, 1985) in order to balance the

right of information access and privacy protection of Canadians. Canada also has the Personal

Information Protection and Electronic Documents Act (PIPEDA) to enforce the use and

collection of personal information in the private sectors (The Office of the Privacy

Commissioner of Canada, 2009). The province of Alberta has Freedom of Information and

Protection of Privacy Act (FOIP) (Alberta Legislature, 2000) and Personal Information

Protection Act (PIPA) (Alberta Legislature, 2003), defining the boundary of limits to which

public bodies (i.e. government) collect, use and disclose of personal information, a framework to

protect the privacy of Albertans for information held within public bodies.

In terms of legislated entities, the Office of the Privacy Commissioner of Canada

(Government of Canada, 2012) and the Office of the Information and Privacy Commissioner of

Page 129: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 129

129

Alberta (OIPC, 2012) represent residents on issues and concerns related to privacy in Canada

and Alberta, respectively.

Respecting privacy of individuals often results in conflict with the intrinsic values of Big

Data. Sensor data that came from surveillance video, mobile phone GPS data, biometrics data

from implanted devices. Other electronic communication data such as email messages, text

messages, digitalized voice and video messages as well as web traffic log. They are all examples

of what constitute Big Data that are advantageous to Predictive Analytics for reasons already

discussed in this essay. However, these data sources suggest an incompatible view of personal

privacy.

The fear of information misuse is one aspect of the overall privacy concern. Take GPS

data as an example, location-aware applications and devices that track the physical location of an

individual. This would bring convenience from photo geotagging to navigation support, but they

also brings the concern of information exploitation. Consider a geotagged picture taken inside of

an individual’s own home which had been inadvertently circulated to the public internet by a

third party. The photo might reveal valuable objects within the home that pique the interest of a

perpetrator. Together with the geotagged information (i.e. latitude and longitude) embedded

within the photo, this will make a compelling case of potential criminal consequence due to

privacy invasion.

As much as data, in and of itself, are morally neutral, the perception of data when view

under different contexts is what challenge researchers to operate within the confines of personal

privacy and professional ethics. Applying Predictive Analytics allows us to make observations

and decisions about the future based on presumptions; that is, applying statistical inference to

predict into the future state based on data. There is always a level of uncertainty with prediction

Page 130: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 130

130

and must not be considered as a guaranteed outcome. For instance, as discussed in Law

Enforcement section of CHAPTER III, predicting recidivism of offenders is an approximation

and not an exact measure. It can be used in supporting parole decisions but certainly cannot be

used in conviction of a future crime. Also, consider the social and moral consequences of job

candidate selection based entirely on predictive scores deduced from the Big Data. Predictive

Analytics application that targets individuals based on their shared attributes inferred by the

individuals’ behavioral data can lead to misclassification, discrimination and damage to

reputation. In addition, preemption could undermine our traditional models of justice, due

process and individual freedoms (Office of The Privacy Commissioner of Canada, 2012). Since

there is no certainty in prediction and only accuracy in measures, we cannot determine future

event as absolute but only as probable outcome.

Conclusion

The challenges remain in developing models that are robust to noise but adaptive to

change. These challenges continue to push the ability of researchers to strike a balance between

model overfitting and model underfitting, as well as to account for the dynamic nature of data we

used in prediction.

The issues, challenges and trends discussed in this essay represent only a subset of the

subject matter, albeit important ones, in understanding the limitations of Predictive Analytics in

the context of Big Data. Certainly, the incorporation of Big Data has been proven to advance to

the application of Predictive Analytics in both theoretical and pragmatic senses. The IBM

Watson Jeopardy! challenge captured the interests of the general public in what Predictive

Analytics with Big Data can achieve. The enablers behind IBM Watson’s ability to answer

natural language questions comprises of multilayers of advanced technologies.

Page 131: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 131

131

Predictive Analytics applies to a wide-ranging applications as previously discussed, novel

means of predictive application will only become more widespread in the near future. The trend

in predictive application is as evidenced by the recent patient filed by Amazon Inc. for the

anticipatory package shipping process (Amazon Inc., 2013).

The fusion of Big Data and ensemble methods employed by IBM Watson marked a

milestone in Predictive Analytics as the Netflix Prize winner team did in exploiting collective

wisdom in deriving predictions. Mining unstructured data with complex data types will continue

to challenge researchers to invent novel means to tackle problems. However, in doing so,

navigating the fine line between accurate predictions and the privacy of individuals, remains a

great challenge. Researchers and practitioners must abide by the ethical principles with

professional attitudes to respect with rights of individuals.

Page 132: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 132

132

CHAPTER VI

CONCLUSIONS AND RECOMMENDATIONS

Conclusions

This essay paper took the position that Predictive Analytics is a specialized interest area

within the Data Mining domain. It focuses on prediction that makes inferences to unknowns and

for that reason, it is considered as a subset of Data Mining field. Big Data and Predictive

Analytics might have begun as marketing terms with the intent to attract attentions and excite

people into the discussion of the subject matter. However, the current landscape is such that the

terms are now synonymous to the ideas, methodologies, methods and techniques that the terms

symbolized.

For these reasons, this essay attempts to elucidate the confusion by starting the discussion

with an overview of the current landscape of the subject matter in CHAPTER I and CHAPTER

II. Then, cross-industry case studies were discussed in CHAPTER III as well as the commonly

employed methodologies were discussed in CHAPTER IV. This lead to CHAPTER V on the

discussion of the issues, challenges and trends of the subject matter. At various points during the

discussion, the juxtaposition between Data Mining and Predictive Analytics was provided and

the symbiotic relationship between Predictive Analytics and Big Data was demonstrated

throughout the essay. Illustrations using figures and tables were created or referenced and cited

to provide visualization in areas that best illustrate the concepts embedded within.

During the course of the construction of this essay as well as the copious amount of

literatures reviewed, a pattern began to emerge in terms of shared practices between the different

practice domains. The result of this realization was captured in the content of this essay as well

as the developed taxonomy diagrams as shown in Figure 1, Figure 3, Figure 6 and Figure 10. The

Page 133: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 133

133

taxonomy diagrams were an attempt to differentiate the constructs in the convoluted field of

analytics, a visualization tool created towards the goal of solidifying the understanding of the

subject matter in the readers’ minds of this essay.

Researchers and practitioners should not blindly believe that Big Data or Predictive

Analytics as a one-size-fits-all solution to our problems. It is however, a solid decision support

tool that further broadens our reach on data that results in the previously unattainable rapid

knowledge acquisition.

Suggestions for Further Research

The torrent of Big Data came with tremendous opportunities and also costs for

materializing the potentials for data analytics. This research indicated that the Big Data evolution

is currently at the early stage, the full potential has yet to be realized. To put this in perspective, a

direct quote from Eric Schmidt, CEO of Google Inc. in year 2010: “There were 5 Exabytes of

information created between the dawn of civilization through 2003, but that much information is

now created every 2 days.”. Almost four years after the statement, we saw the advancement of

NoSQL databases, Cloud Computing and the ecosystems that are built around open source

platforms such as OpenStack and Hadoop.

SOA

From a software architectural perspective, we have NoSQL databases and Hadoop

platform to handle the challenges brought by Big Data. From an infrastructural perspective, we

have Cloud Computing platform (i.e. IaaS, PaaS and SaaS) with hardware visualization to bring

operating cost down (i.e. utility computing) to support the many applications of Predictive

Analytics with Big Data. To that end, the infrastructure and the toolsets are now in place to

handle Big Data. The industry as a whole will continue to evolve around the Software Oriented

Page 134: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 134

134

Architecture (SOA) to bring Data-as-a-Service (DaaS) to a more mainstream level, especially in

the context of Big Data.

Predictive Analytics will also capitalize on SOA to bring organizations the services of

Predictive-Analytics-as-a-Service (PAaaS). Combining DaaS and PAaaS, the end result of the

full realization and ubiquitous adoption of this synergized technology, is what we can look

forward to in the next frontier in the application of Predictive Analytics in Data Mining with Big

Data.

With that said, the research community thus needs to focus more on the SOA aspect

rather than on the modeling methods and algorithms advancements, as we have reached a plateau

in those dimensions as discussed in previous chapters. Therefore, our collective attentions should

concentrate on making the technologies more available, accessible and affordable to the

community at large.

Real-time analytics

Another challenge related to Big Data and Predictive Analytics is the ability to process

information at real-time or at nearly real-time speed. The early version of the Hadoop Platform

did not provide adequate support on this front as the MapReduce method (Vavilapalli, et al.)

operates in batch processing mode. Under the batch processing mode, the notion of a job is

intrinsically a scheduling-based activity, which is to say, a job based approach is incompatible

with real-time application as there always exists a significant delay between job executions.

In addressing these issues with real-time data processing, the Hadoop Platform version

2.0 introduced a number of components including Hadoop YARN and Apache STORM to

provide continuous real-time streaming analytics. Other important Apache projects that deal with

real-time Big Data problems include Apache Spark, Apache Drill and Apache S4. Reducing the

Page 135: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 135

135

time gap between recording an observation and making a decision will maximize the effect of

any prediction.

Making timely decision is critical to the application of Predictive Analytics. However,

even with the accompanying technologies introduced in Hadoop 2.0, the exponential growth of

Big Data would eventually pose a risk in spite of the most cutting-edge developments in the

Apache Hadoop project. There exists an asymmetrical growth rate between Big Data and

Hadoop. Big Data is projected to grow exponential until the year 2020 (Gantz, Reinsel, & Lee,

2013) while the current Hadoop architecture currently scale linearly. The gap will continue to

grow between data that are available and the ability to process them in the timely manner.

NoSQL

The NoSQL data model offers the following characteristics: Basically Available, Soft

State, and Eventual Consistency. These characteristics are commonly known as BASE. Contrary

to ACID (Atomicity, Consistency, Isolation and Durability), which is common amongst relational

data model, BASE offers a less rigid data model that favors performance over strong data

consistency.

The reason that many adopted the NoSQL data model for Big Data is because, NoSQL

data model is able to handle the characteristics of Big Data where relational data model fall

short. The shortcomings in NoSQL data model as it relates to relational data model becomes the

strength in which NoSQL can excel in an era of Big Data.

If ensuring data atomicity and strong data consistency are too difficult to do with Big

Data and relational data model, then we would design our applications around these limitations

and learn to accept the trade-offs using NoSQL. However, there are situations where we would

expect strong transactionality for our operations such as securities trading data and financial

Page 136: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 136

136

transactions. These situations can also be benefited by Big Data and thus, an improved version of

NoSQL with better transactionality support is desirable. Solutions such as FoundationDB

(FoundationDB, 2014) have made the first step in improving NoSQL with ACID support. More

research is needed to validate the viability and to determine the cost of ACID based NoSQL data

model. To that end, a hybridized data model is to be looked forward to that fuses the benefits

between the relational data model and the NoSQL data model.

Page 137: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 137

137

REFERENCES

About Us Background. (2014, 01 29). Retrieved from European Molecular Biology Laboratory -

European Bioinformatics Institute: http://www.ebi.ac.uk/about/background

Academia.edu. (2014, February). Retrieved from Academia.edu: http://www.academia.edu/

academicindex.net. (2014, February). Retrieved from AcademicIndex.net:

http://www.academicindex.net/

ACM Digital Library. (2014, February). Retrieved from ACM.org: http://dl.acm.org/

Alberta Legislature. (2000). Freedom of Information and Protection of Privacy Act. Retrieved

from Service Alberta - Queen's Printer:

http://www.qp.alberta.ca/1266.cfm?page=F25.cfm&leg_type=Acts&isbncln=978077976

2071

Alberta Legislature. (2003). Personal Information Protection Act. Retrieved from Service

Alberta > Queen's Printer:

http://www.qp.alberta.ca/1266.cfm?page=P06P5.cfm&leg_type=Acts&isbncln=9780779

762507

Amazon Inc. (2013, December). Patent #: US008615473 Section: Specifications 14 of 27 pages .

Retrieved from United States Patent and Trademark Office:

http://pdfpiw.uspto.gov/.piw?docid=08615473&SectionNum=3&IDKey=2809BD982F3

D&HomeUrl=http://patft.uspto.gov/netacgi/nph-

Parser?Sect1=PTO2%2526Sect2=HITOFF%2526p=1%2526u=%25252Fnetahtml%2525

2FPTO%25252Fsearch-bool.html%2526r=1%2526f=G%2526l=50%2526co1=AND%2

Anderson, P. W. (1972). More Is Different. Retrieved from

http://www.ph.utexas.edu/~wktse/Welcome_files/More_Is_Different_Phil_Anderson.pdf

Page 138: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 138

138

Annika Wolff, Z. Z. (2013). Improving retention: predicting at-risk students by analysing

clicking behaviour in a virtual learning environment. LAK '13: Proceedings of the Third

International Conference on Learning Analytics and Knowledge (pp. 145-149). ACM.

doi:10.1145/2460296.2460324

ARIAS, M., ARRATIA, A., & XURIGUERA, R. (2013, December). Forecasting with Twitter

Data. ACM Transactions on Intelligent Systems and Technology, 5, pp. 8:1-8:24.

doi:10.1145/2542182.2542190

Athabasca University Library. (2014, Feb). Retrieved from Athabasca University:

http://library.athabascau.ca/

Ayhan, S., Pesce, J., Comitz, P., Gerberick, G., & Bliesner, S. (2012). Predictive analytics with

surveillance big data. BigSpatial '12 Proceedings of the 1st ACM SIGSPATIAL

International Workshop on Analytics for Big Geospatial Data (pp. 81-90). ACM.

Baldauf, M., Dustdar, S., & Rosenberg, F. (2007). A survey on context-aware systems;Schahram

Dustdar;Florian Rosenberg. Int. J. Ad Hoc and Ubiquitous Computing, 2. Retrieved from

http://www.cs.umd.edu/class/spring2013/cmsc818b/files/surveyoncontextawaresystems.p

df

Barber, R., & Sharkey, M. (2012). Course Correction: Using Analytics to Predict Course

Success. LAK '12: Proceedings of the 2nd International Conference on Learning

Analytics and Knowledge (pp. 259-262). ACM. doi:10.1145/2330601.2330664

Begemana, K., Belikov, A., Boxhoorna, D., Dijkstrab, F., Holties, H., Meyer-Zhaob, Z., . . .

Vrienda, W.-J. (2010, August 21). LOFAR Information System. Future Generation

Computer Systems, pp. 319-328.

Page 139: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 139

139

Berndt, D. J., Bhat, S., Fisher, J. W., Hevner, A. R., & Studnicki, J. (2004). Data Analytics for

Bioterrorism Surveillance. Intelligence and Security Informatics, 3073, 17-27.

Bigus, J. P., Chitnis, U., Deshpande, P. M., Kannan, R., Mohania, M. K., Negi, S., . . . White, B.

F. (2009). CRM Analytics Framework. 15th International Conference on Management of

Data. India. Retrieved from

http://www.cse.iitb.ac.in/~comad/2009/proceedings/R2_2.pdf

CBC. (n.d.). Google helps predict flu case surge for hospitals. Retrieved from CBC News:

http://www.cbc.ca/news/technology/story/2012/01/10/sci-google-flu-hospitals.html

Ceri, S., Valle, E. D., Pedreschi, D., & Trasarti, R. (2012). Mega-modeling for Big Data

Analytics. Conceptual Modeling, 7532, 1-15.

Clair, M. S. (2011). So Much, So Fast, So Little Time.

Codd, E. F. (1970, June). A relational model of data for large shared data banks. Communications

of the ACM, 13(6). doi:10.1145/362384.362685

Colwiz. (2014, February). Retrieved from Colwiz.com: http://www.colwiz.com/

Computing. (2014, 01 29). Retrieved from CERN: http://home.web.cern.ch/about/computing

Couchbase.com. (2012). Couchbase Server Technical Overview. Retrieved from

http://www.couchbase.com/sites/default/files/uploads/all/whitepapers/Couchbase-Server-

Technical-Whitepaper.pdf

Das, K. K., Fratkin, E., Gorajek, A., Stathatos, K., & Gajjar, M. (2011). Massively parallel in-

database predictions using PMML. PMML '11 Proceedings of the 2011 workshop on

Predictive markup language modeling (pp. 22-27). ACM. doi:10.1145/2023598.2023601

Data Mining Group. (2014). Home. Retrieved from Data Mining Group: http://www.dmg.org/

Page 140: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 140

140

DeCandia, G., Hastorun, D., Jampani, M., Kakulapati, G., Lakshman, A., Pilchin, A., . . . Vogels,

W. (2007). Dynamo: Amazon’s Highly Available Key-value Store. Retrieved from

Dynamo: Amazon’s Highly Available Key-value Store

Dey, A. K. (2001). Understanding and Using Context. Future Computing Environments Group.

Georgia Institute of Technology. doi:10.1007/s007790170019

Digital Library of the Commons Repository. (2014, February). Retrieved from Indiana

University: http://dlc.dlib.indiana.edu/dlc

DMG. (2014). PMML 4.1 - General Structure. Retrieved from Data Mining Group (DMG):

http://www.dmg.org/v4-1/GeneralStructure.html

dogpile. (2014, February). Retrieved from Dogpile.com: http://www.dogpile.com/

EBSCO Colleges and Universities Online Resources for Academic Libraries. (2014, February).

Retrieved from EBSCO: http://www.ebscohost.com/academic

EMC Inc. (2012). Smart Grid Analytics on EMC Greenplum in Partnership with Silver Spring

Networks. Retrieved from http://www.emc.com/collateral/hardware/white-papers/h8762-

smart-grid-analytics-greenplum-ssn.pdf

Evernote. (2014, February). Retrieved from Evernote.com: https://evernote.com/

Financial Services Authority (FSA) UK. (2006). Markets Division: Newsletter on Market

Conduct and Transaction Reporting Issue - Alpha Capture Systems. FSA. Retrieved from

http://www.fsa.gov.uk/pubs/newsletters/mw_newsletter17.pdf

Fischer, U., Dannecker, L., Siksnys, L., Rosenthal, F., Boehm, M., & Lehner, W. (2013, March).

Towards Integrated Data Analytics: Time Series Forecasting. Datenbank-Spektrum,

13(1), 45-53. doi:10.1007/s13222-012-0108-4

FoundationDB. (2014). Retrieved from FoundationDB: https://foundationdb.com/

Page 141: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 141

141

Freund, Y., & Schapire, R. E. (1999, September). A Short Introduction to Boosting. Journal of

Japanese Society for Artificial Intelligence, 771-780. Retrieved from

http://cseweb.ucsd.edu/~yfreund/papers/IntroToBoosting.pdf

Fülöp, L. J., Beszédes, Á., Tóth, G., Demeter, H., Vidács, L., & Farkas, L. (2012). Predictive

complex event processing: a conceptual framework for combining complex event

processing and predictive analytics. BCI '12: Proceedings of the Fifth Balkan Conference

in Informatics (pp. 26-31). ACM. doi:10.1145/2371316.2371323

Gantz, J., Reinsel, D., & Lee, R. (2013). THE DIGITAL UNIVERSE IN 2020: Big Data, Bigger

Digi tal Shadows, and Biggest Growth in the Far East. EMC Corporation. Retrieved

from http://www.emc.com/collateral/analyst-reports/emc-digital-universe-china-brief.pdf

Gartner. (2013). Magic Quadrant for Operational Database Management Systems. Gartner.

Retrieved from http://www.gartner.com/technology/reprints.do?id=1-

1M9YEHW&ct=131028&st=sb

Gartner. (2014). Magic Quadrant for Business Intelligence and Analytics Platforms. Gartner.

Retrieved from http://www.gartner.com/technology/reprints.do?id=1-

1QLGACN&ct=140210&st=sb

Gartner.com. (2014, 06 30). Gartner IT Glossary - Big Data. Retrieved from Gartner.com:

http://www.gartner.com/it-glossary/big-data/

Geisler, B. (2002). An Empirical Study of Machine Learning Algorithms Applied to Modeling

Player Behavior in a ìFirst Person Shooterî Video Game. Computer Sciences.

UNIVERSITY OF WISCONSIN - MADISON. Retrieved from

http://ftp.cs.wisc.edu/machine-learning/shavlik-group/geisler.thesis.pdf

GNU PSPP. (2014, February). Retrieved from gnu.org: http://www.gnu.org/software/pspp/

Page 142: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 142

142

Google. (2014, February). Retrieved from Google.com: http://www.google.com

Google Correlate. (2014, February). Retrieved from Google.com:

http://www.google.com/trends/correlate/

Google Inc. (n.d.). Google Flu Trends. Retrieved from How does it work?:

http://www.google.org/flutrends/about/how.html

Google Scholar. (2014, February). Retrieved from Google.com: http://scholar.google.ca/

Google Trends. (2014, February). Retrieved from Google.com: http://www.google.ca/trends/

Government of Canada. (1985). Access to Information Act. Retrieved from Justice Laws

Website: http://laws-lois.justice.gc.ca/eng/acts/A-1/FullText.html

Government of Canada. (1985). Privacy Act. Retrieved from Justice Laws Website: http://laws-

lois.justice.gc.ca/eng/acts/P-21/FullText.html

Government of Canada. (2012, December 5). About the Office of the Privacy Commissioner.

Retrieved 2014, from Office of the Privacy Commissioner of Canada:

http://www.priv.gc.ca/au-ans/index_e.asp

Greengard, S. (2012, March). Policing the Future. Communications of the ACM, 19-21.

doi:10.1145/2093548.2093555

Grimes, S. (2012). Frequently Asked Questions about In-Database Analytics. Retrieved from

http://www.teradata.com/white-papers/Frequently-Asked-Questions-about-In-Database-

Analytics-eb6189/?type=WP

Guazzelli, A., Jena, T., Lin, W.-C., & Zeller, M. (2011). The PMML Path towards True

Interoperability in Data Mining. PMML '11 Proceedings of the 2011 workshop on

Predictive markup language modeling (pp. 32-38). ACM. doi:10.1145/2023598.2023603

Page 143: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 143

143

Guazzelli, A., Stathatos, K., & Zeller, M. (2009, Nov). Efficient Deployment of Predictive

Analytics through Open Standards and Cloud Computing. SIGKDD Explorations

Newsletter, pp. 32-38. doi:10.1145/1656274.1656281

Haas, P. J., Maglio, P. P., Selinger, P. G., & Tan, a. W.-C. (2011). Data is Dead… Without What-

If Models. Retrieved from http://users.soe.ucsc.edu/~wctan/papers/2011/splash-vldb.pdf

Han, J., Kamber, M., & Pei, J. (2011). In J. Han, M. Kamber, & J. Pei, Data Mining: Concepts

and Techniques, Third Edition. Morgan Kaufmann. doi:978-0123814791

Han, J., Kamber, M., & Pei, J. (2011). In J. Han, M. Kamber, & J. Pei, Data Mining: Concepts

and Techniques, Third Edition. Morgan Kaufmann. doi:978-0123814791

Hardoop. (2014, March 7). Welcome to Apache Pig! Retrieved from Hardoop:

https://pig.apache.org/

HDFS Architecture Guide. (2014, 05 11). Retrieved from Apache Hadoop:

http://hadoop.apache.org/docs/r1.2.1/hdfs_design.html

Hollywood, J., Smith, S., Price, C., McInnis, B., & Perry, W. (2012). Predictive Policing: What It

Is, What It Isn't, and Where It Can Be Useful. Retrieved from International Association of

Chiefs of Police: http://www.theiacp.org/Portals/0/pdfs/LEIM/2012Presentations/OPS-

PredictivePolicing.pdf

IBM. (2014, 01 26). Importing and Exporting Models as PMML. Retrieved from IBM:

http://pic.dhe.ibm.com/infocenter/spssmodl/v15r0m0/index.jsp?topic=%2Fcom.ibm.spss.

modeler.help%2Fmodels_import_pmml.htm

IBM Inc. (2012). IBM Netezza Analytics. Retrieved from

http://public.dhe.ibm.com/common/ssi/ecm/en/imd14365usen/IMD14365USEN.PDF

Page 144: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 144

144

IBM Inc. (2013). WellPoint, Inc. IBM. Retrieved from http://www-

03.ibm.com/innovation/us/watson/pdf/WellPoint_Case_Study_IMC14792.pdf

IEEE Xplore Digital Library. (2014, February). Retrieved from IEEE.org:

http://ieeexplore.ieee.org/Xplore/home.jsp

IET Inspec. (2014, Feb). Retrieved from The Institution of Engineering and Technology:

http://www.theiet.org/resources/inspec/

IPython. (2014, February). Retrieved from IPython.org: http://ipython.org/

iSeek Education. (2014, February). Retrieved from iSeek.com: http://education.iseek.com

James W. Pennebaker; Roger J. Booth; Martha E. Francis. (2014, March 8). Linguistic Inquiry

and Word Count. Retrieved from http://www.liwc.net/: http://www.liwc.net/

Jarrold, W. L., Peintner, B., Yeh, E., Krasnow, R., Javitz, H. S., & Swan, G. E. (2010). Language

Analytics for Assessing Brain Health: Cognitive Impairment, Depression and Pre-

symptomatic Alzheimer’s Disease. Brain Informatics, 6334, 299-307.

Jaskowski, M., & Jaroszewicz, S. (2012). Uplift modeling for clinical trial data. ICML 2012

Workshop on Clinical Data Analysis, Edinburgh, Scotland, UK, 2012. Retrieved from

http://people.cs.pitt.edu/~milos/icml_clinicaldata_2012/Papers/Oral_Jaroszewitz_ICML_

Clinical_2012.pdf

Jennings, W. G., & M.C.J. (2006). Revisiting Prediction Models in Policing: Identifying High-

Risk Offenders. AMERICAN JOURNAL OF CRIMINAL JUSTICE.

doi:10.1007/BF02885683

Kaur, B., Rawat, S., Dinesh, R., Ghosh, S., Puri, M., Das, A., & Sengar, J. S. (2013). A Novel

Approach to 8-Queen Problem Employing Machine Learning.

Page 145: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 145

145

KhanAcademy Probability and Statistics. (2014, February). Retrieved from KhanAcademy.org:

https://www.khanacademy.org/math/probability

Kimbrough, S. O., Chou, C., Chen, Y.-T., & Lin, H. (2013, August). On developing indicators

with text analytics: exploring concept vectors applied to English and Chinese texts.

Information Systems and e-Business Management. doi:10.1007/s10257-013-0228-x

Kiseleva, J. (2013). Context mining and integration into predictive web analytics. WWW '13

Companion: Proceedings of the 22nd international conference on World Wide Web

companion (pp. 383-387). International World Wide Web Conferences Steering

Committee.

KNIME. (2014, February). Retrieved from KNIME - Professional Open-Source Software:

http://www.knime.org/

KNIME. (2014, 01 26). Export and Convert R models to PMML within KNIME. Retrieved from

KNIME: http://www.knime.org/blog/export-and-convert-r-models-pmml-within-knime

Koren, Y. (2009). The BellKor Solution to the Netflix Grand Prize. Retrieved from

http://www.netflixprize.com/assets/GrandPrize2009_BPC_BellKor.pdf

Kridel, D., & Dolk, D. (2013). Automated self-service modeling: predictive analytics as a

service. Information Systems and e-Business Management, 11(1), 119-140.

doi:10.1007/s10257-011-0185-1

Lazer, D., Kennedy, R., King, G., & Vespignani, A. (2014). The Parable of Google Flu: Traps in

Big Data Analysis. Science. doi:10.1126/science.1248506

Leiner, B. M., Cerf, V. G., Clark, D. D., Kahn, R. E., Kleinrock, L., Lynch, D. C., . . . Wolff, S.

(n.d.). Brief History of the Internet. Retrieved from Internet Society:

Page 146: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 146

146

http://www.internetsociety.org/internet/what-internet/history-internet/brief-history-

internet

Lesk, M. (n.d.). How Much Information Is There In the World? Retrieved from Lesk.com:

http://www.lesk.com/mlesk/ksg97/ksg.html

Lesson 19: Logistic Regression. (2014). Retrieved from Professor Courtney Brown's YouTube

Video Classes:

http://www.courtneybrown.com/classes/video_classes_Courtney_Brown.html

Lin, J., & Kolcz, A. (2012). Large-Scale Machine Learning at Twitter. SIGMOD '12:

Proceedings of the 2012 ACM SIGMOD International Conference on Management of

Data (pp. 793-804). ACM. doi:10.1145/2213836.2213958

Lin, J., & Ryaboy, D. (2013). Scaling big data mining infrastructure: the twitter experience.

SIGKDD Explorations Newsletter, 6-19.

LKI: Environment for Developing KDD-Applications Supported by Index-Structures. (2014,

February). Retrieved from LMU: http://elki.dbs.ifi.lmu.de/

Llaves, A., & Maué, P. (2011). Processing Events in an Affective Environment.

Machine Learning. (2014, March 10). Retrieved from Dlib C++ Library: http://dlib.net/ml.html

Maciejewski, R., Hafen, R., Rudolph, S., Larew, S. G., Mitchell, M. A., Cleveland, W. S., &

Ebert, D. S. (2011). Forecasting Hotspots—A Predictive Analytics Approach. IEEE

Computer Society. doi:1077-2626

Martin, P., Matheson, M., Lo, J., Ng, J., Tan, D., & Thomson, B. (2010). Supporting Smart

Interactions with Predictive Analytics. The Smart Internet, 6400, 103-114.

Page 147: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 147

147

Masud, M. M., Gao, J., Khan, L., Han, J., & Thuraisingham, B. (2009). Integrating Novel Class

Detection with Classification for Concept-Drifting Data Streams. Springer-Verlag Berlin

Heidelberg. Retrieved from http://www.cs.uiuc.edu/~hanj/pdf/pkdd09_mmasud.pdf

Mathewos, B., Carvalho, M., & Ham, F. (2011). Network traffic classification using a parallel

neural network classifier architecture. CSIIRW '11 Proceedings of the Seventh Annual

Workshop on Cyber Security and Information Intelligence Research. New York: ACM.

doi:10.1145/2179298.2179334

Matplotlib. (2014, Feb). Retrieved from matplotlib.org: http://matplotlib.org/

Mendeley. (2014, February). Retrieved from Mendeley.com: http://www.mendeley.com/

Microsoft Academic Research. (2014, Feb). Retrieved from Microsoft.com:

http://academic.research.microsoft.com/

Moore, H. (2013, September 13). Twitter heads for stock market debut by filing for IPO.

Retrieved from The Guardian:

http://www.theguardian.com/technology/2013/sep/12/twitter-ipo-stock-market-

launch?CMP=EMCNEWEML6619I2&et_cid=48826&et_rid=7107573&Linkid=http%3a

%2f%2fwww.theguardian.com%2ftechnology%2f2013%2fsep%2f12%2ftwitter-ipo-

stock-market-launch

mpmath. (2014, Feb). Retrieved from code.google.com: https://code.google.com/p/mpmath/

Murray, B. H., & Moore, A. (2000, July 10). Sizing The Internet. Retrieved from University of

Toronto: http://www.cs.toronto.edu/~leehyun/papers/Sizing_the_Internet.pdf

Nankani, E., & Simoff, S. (n.d.). Predictive analytics that takes in account network relations: A

case study of research data of a contemporary university. AusDM '09: Proceedings of the

Eighth Australasian Data Mining Conference - Volume 101 (pp. 99-108). ACM.

Page 148: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 148

148

Natural Resources Canada. (2012, 10 18). Retrieved from Climate and Climate-related Trends

and Projections: http://www.nrcan.gc.ca/environment/resources/publications/impacts-

adaptation/reports/assessments/2008/10261

Nauck, D. D., Ruta, D., Spott, M., & Azvine, B. (2006). Being proactive – analytics for

predicting customer actions. BT Technology Journal, 24(1), 17-26.

Niculescu-Mizil, A., & Caruana, R. (2005). Predicting Good Probabilities With Supervised

Learning. Proceedings of the 22 nd International Conference. Cornell University.

Retrieved from

http://machinelearning.wustl.edu/mlpapers/paper_files/icml2005_Niculescu-

MizilC05.pdf

Nirkhi, S. (2010). Potential use of Artificial Neural Network in Data Mining. G.H.Raisoni

College of Engineering, Department of Computer Science. India: IEEE. doi:978-1-4244-

5586-7

NumPy. (2014, Feb). Retrieved from Numpy.org: http://www.numpy.org/

Nyce, C. (2007). Predictive Analytics White Paper. American Institute for CPCU/Insurance

Institute of America. Retrieved from

http://www.theinstitutes.org/doc/predictivemodelingwhitepaper.pdf

Office of The Privacy Commissioner of Canada. (2012, August). The Age of Predictive

Analytics: From Patterns to Predictions. Retrieved from Privacy Research Papers:

http://www.priv.gc.ca/information/research-recherche/2012/pa_201208_e.asp

OIPC. (2012). Office. Retrieved from Office of the Information and Privacy Commissioner of

Alberta: http://www.oipc.ab.ca/pages/About/Office.aspx

Page 149: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 149

149

Oracle Inc. (2011). In-Database Analytics: Predictive Analytics, Data Mining, Exadata &

Business Intelligence. Retrieved from

http://www.oracle.com/technetwork/database/options/advanced-analytics/odm/oracle-in-

database-analytics-oow11-517499.pdf

Oracle Inc. (2013). Big Data Analytics. Retrieved from

http://www.oracle.com/technetwork/database/options/advanced-analytics/advanced-

analytics-wp-12c-1896138.pdf?ssSourceSiteId=ocomen

Orange 2.7 for Windows. (2014, February). Retrieved from Orange: http://orange.biolab.si/

pandas. (2014, Feb). Retrieved from pandas.pydata.org: http://pandas.pydata.org/

Patterns of Inference. (2014, 2 23). Retrieved from Probabilistic Models of Cognition:

https://probmods.org/patterns-of-inference.html

Pearl, J., & Russell, S. (2000). Bayesian Networks. University of California. Retrieved from

http://www.cs.berkeley.edu/~russell/papers/hbtnn-bn.pdf

Piatetsky-Shapiro, G. (2007, August). Data mining and knowledge discovery 1996 to 2005:

overcoming the hype and moving from “university” to “business” and “analytics”. Data

Mining and Knowledge Discovery, 15(1), 99-105. doi:10.1007/s10618-006-0058-2

Preis, T., Moat, H. S., & Stanley, H. E. (n.d.). Quantifying Trading Behavior in Financial

Markets Using Google Trends. Retrieved from

http://www.nature.com/srep/2013/130425/srep01684/full/srep01684.html

Qiqqa. (2014, February). Retrieved from Qiqqa.com: http://www.qiqqa.com/

R. (2014, 01 26). Generate PMML for various models. Retrieved from R: http://cran.r-

project.org/web/packages/pmml/index.html

Page 150: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 150

150

RapidMiner. (2014, 01 26). RapidMiner Extensions PMML. Retrieved from SourceForge:

http://sourceforge.net/projects/rapidminer/files/2.%20Extensions/PMML/5.0/

RapidMiner STARTER Edition. (2014, February). Retrieved from RapidMiner:

http://rapidminer.com/

refseek. (2014, February). Retrieved from RefSeek.com: http://www.refseek.com/

Riensche, R. M., & Whitney, P. D. (2012, August). Combining modeling and gaming for

predictive analytics. Security Informatics, 1(11).

Rzepakowski, P., & Jaroszewicz, S. (2011). Decision trees for uplift modeling with single and

multiple treatments. Poland: Springer. doi:10.1007/s10115-011-0434-0

Sanfilippo, A., Butner, S., Cowell, A., Dalton, A., Haack, J., Kreyling, S., . . . Whitney, P. (2011).

Technosocial Predictive Analytics for Illicit Nuclear Trafficking. Social Computing,

Behavioral-Cultural Modeling and Prediction, 6589, 374-381.

SAS Inc. (2007). SAS In-Database Processing. Retrieved from

http://support.sas.com/resources/papers/InDatabase07.pdf

Schwegmann, B., Matzner, M., & Janiesch, C. (2013). preCEP: Facilitating Predictive Event-

Driven Process Analytics. Design Science at the Intersection of Physical and Virtual

Design, 7939, 448-455.

scikit-image: Image Processing in Python. (2014, February). Retrieved from scikit-image.org:

http://scikit-image.org/

scikit-learn: Machine Learning in Python. (2014, February). Retrieved from scikit-learn.org:

http://scikit-learn.org/

SciPy Stack. (2014, February). Retrieved from SciPy.org: http://www.scipy.org/about.html

Page 151: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 151

151

Shimpi, D., & Chaudhari, S. (2012). An overview of Graph Databases. International Conference

in Recent Trends in Information Technology and Computer Science (ICRTITCS - 2012

(pp. 16-22). International Journal of Computer Applications® (IJCA) (0975 – 8887).

Retrieved from http://research.ijcaonline.org/icrtitcs2012/number3/icrtitcs1351.pdf

Shmueli, G. (2010). To Explain or To Predict? Statistical Science. Retrieved from

http://papers.ssrn.com/sol3/papers.cfm?abstract_id=1351252

Shmueli, G., & Koppius, O. (2010). Predictive Analytics in Information Systems Research.

doi:http://papers.ssrn.com/sol3/papers.cfm?abstract_id=1606674

Short, M. B., D’Orsogna, M. R., Brantingham, P. J., & Tita, G. E. (2009). Measuring repeat and

near-repeat burglary effects. doi:10.1007/s10940-009-9068-8

Siegel, E. (2013). Predictive Analytics: The Power to Predict Who Will Click, Buy, Lie, or Die.

Wiley.

Singh, N., & Rao, S. (2013). Meta-Learning based Architectural and Algorithmic Optimization

for Achieving Green-ness in Predictive Workload Analytics. SAC '13 Proceedings of the

28th Annual ACM Symposium on Applied Computing (pp. 1169-1176). ACM.

doi:10.1145/2480362.2480582

Soulas, E., & Shasha, D. (2013). Online Machine Learning Algorithms For Currency Exchange

Prediction. NYU, Courant Department. New York: NYU CS. Retrieved from

http://cs.nyu.edu/web/Research/TechReports/TR2013-953/TR2013-953.pdf

SpringerLink. (2014, February). Retrieved from Springer.com: http://link.springer.com/

Statsmodels. (2014, February). Retrieved from Sourceforget.net:

http://statsmodels.sourceforge.net/

Page 152: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 152

152

StatSoft. (2014, March 10). STATISTICA Product Index. Retrieved from StatSoft:

http://www.statsoft.com/Products/STATISTICA/Product-Index

SymPy. (2014, February). Retrieved from sympy.org: http://sympy.org/

Teradata . (2013, Sept 18). Retrieved from Teradata Offers First, Fully-Parallel, Scalable R

Analytics: http://www.teradata.com/News-Releases/2013/Teradata-Offers-First-Fully-

Parallel-Scalable-R-Analytics/

The Office of the Privacy Commissioner of Canada. (2009, April). A Guide for Individuals - Your

Guide to PIPEDA. Retrieved from Office of The Privacy Commissioner of Canada: The

Personal Information Protection and Electronic Documents Act

The R Project for Statistical Computing. (2014, February). Retrieved from R: http://www.r-

project.org/

Thomas, J. W. (2011, January). Capturing Alpha in the Alpha Capture System: Do Trade Ideas

Generate Alpha? The Journal of Investing, 20(1), 11-18. Retrieved from http://0-

search.ebscohost.com.aupac.lib.athabascau.ca/login.aspx?direct=true&db=edszbw&AN=

EDSZBW670042889&site=eds-live

UCI Machine Learning Repository. (2014, 01 26). Retrieved from UCI Machine Learning

Repository: http://archive.ics.uci.edu/ml/datasets.html

Vavilapalli, V. K., Murthy, A. C., Douglas, C., Agarwal, S., Konar, M., Evans, R., . . .

Baldeschwieler, E. (n.d.). Apache Hadoop YARN: Yet Another Resource Negotiator.

Venkatesan, A., Krishnan, N. C., & Panchanathan, S. (2010). Cost-sensitive Boosting for Concept

Drift.

Virtual LRC. (2014, February). Retrieved from VirtualLRC.com: http://www.virtuallrc.com/

Page 153: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 153

153

Vlachos, M., Domeniconi, C., Gunopulos, D., Kollios, G., & Koudas, N. (2002). Non-Linear

Dimensionality Reduction Techniques for Classification and Visualization. SIGKDD.

Edmonton: ACM. doi:1-58113-567-X/02/0007

Vouk, M. A. (2008). Cloud Computing – Issues, Research and Implementations. Journal of

Computing and Information Technology, 235-246. doi:10.2498/cit.1001391

Waller, M. A., & Fawcett, S. E. (2013). Data Science, Predictive Analytics, and Big Data: A

Revolution That Will Transform Supply Chain Design and Management. Journal of

Business Logistics, 77-84. doi:10.1111/jbl.12010

Wang, D., Navathe, S. B., Liu, L., Irani, D., Tamersoy, A., & Pu, C. (2013). Click Traffic

Analysis of Short URL Spam on Twitter. Collaborative Computing: Networking,

Applications and Worksharing (Collaboratecom), 2013 9th International Conference

Conference (pp. 250-259). IEEE.

Weka. (2014, 01 26). Weka Documentation. Retrieved from Weka - The University of Waikato:

http://www.cs.waikato.ac.nz/ml/weka/documentation.html

Weka 3: Data Mining Software in Java. (2014, February). Retrieved from WEKA:

http://www.cs.waikato.ac.nz/ml/weka/index.html

Welcome to Apache™ Hadoop®. (2014, 05 11). Retrieved from Apache Hadoop:

http://hadoop.apache.org/

What is Apache Mahout? (2014, 03 10). Retrieved from Apache Mahout:

http://mahout.apache.org/

WolframAlpha. (2014, February). Retrieved from WolframAlpha.com:

http://www.wolframalpha.com/

Page 154: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 154

154

Yang, M., Wong, S. C., & Coid, J. (2010). The Efficacy of Violence Prediction: A Meta-Analytic

Comparison of Nine Risk Assessment Tools. American Psychological Association, (pp.

740-766). doi:10.1037/a0020473

Ye, J., Chow, J.-H., Chen, J., & Zheng, Z. (2009). Stochastic Gradient Boosted Distributed

Decision Trees. CIKM’09, (pp. 2061-2064). Retrieved from

http://www.cslu.ogi.edu/~zak/cs506-pslc/sgradboostedtrees.pdf

Zementis. (n.d.). PMML in Action: Data Transformations. Retrieved from Zementis.com:

http://www.zementis.com/PMMLTransformations/PMMLTransformations.html

Zeng, A., & Huang, Y. (2011). A text classification algorithm based on rocchio and hierarchical

clustering. ICIC'11 Proceedings of the 7th international conference on Advanced

Intelligent Computing (pp. 432-439). Berlin: ACM. doi:10.1007/978-3-642-24728-6_59

Zotero. (2014, February). Retrieved from zotero.org: zotero.org

Page 155: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 155

155

APPENDIX A – PMML CODE

PMML CODE EXAMPLE

<DerivedField name="Field2" optype="continuous" dataType="double">

<NormContinuous field="Field1" mapMissingTo="Field3" outliers="asExtremeValues">

<LinearNorm orig="Original Value 1" norm="Normalized Value 1"/>

<LinearNorm orig="Original Value 2" norm="Normalized Value 2"/>

<LinearNorm orig="Original Value 3" norm="Normalized Value 3"/>

</NormContinuous>

</DerivedField>

PMML CODE EXAMPLE - HEADER SECTION

<Header copyright="KNIME">

<Application name="KNIME" version="2.8.0"/>

</Header>

PMML CODE EXAMPLE - DATADICTIONARY SECTION

<DataDictionary numberOfFields="10">

<DataField dataType="integer" name="Age" optype="continuous">

<Interval closure="closedClosed" leftMargin="17.0" rightMargin="90.0"/>

</DataField>

<DataField dataType="string" name="Employment" optype="categorical">

<Value value="Private"/>

<Value value="Consultant"/>

<Value value="SelfEmp"/>

</DataField>

<DataField dataType="string" name="Education" optype="categorical">

<Value value="College"/>

</DataField>

<DataDictionary>

PMML CODE EXAMPLE - TRANSFORMATIONDICTIONARY SECTION

<MapValues outputColumn="longForm">

<FieldColumnPair field="gender" column="shortForm"/>

<InlineTable>

Page 156: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 156

156

<row><shortForm>m</shortForm><longForm>male</longForm>

</row>

<row><shortForm>f</shortForm><longForm>female</longForm>

</row>

</InlineTable>

</MapValues>

PMML CODE EXAMPLE - MODEL SECTION – SUPPORT VECTOR MACHINE

<SupportVectorMachineModel modelName="SVM" functionName="classification" algorithmName="Sequential Minimal

Optimization (SMO)" svmRepresentation="SupportVectors">

<MiningSchema>

<MiningField name="Age" invalidValueTreatment="asIs"/>

<MiningField name="Income" invalidValueTreatment="asIs"/>

...

</MiningSchema>

<Targets>

<Target field="TARGET_Adjusted" optype="categorical">

<TargetValue value="0"/>

<TargetValue value="1"/>

</Target>

</Targets>

<LocalTransformations>

<DerivedField dataType="integer" name="Private_Employment" optype="ordinal">

<NormDiscrete field="Employment" mapMissingTo="0.0" value="Private"/>

</DerivedField>

...

<DerivedField dataType="string" name="Age_binned" optype="categorical">

<Discretize field="Age">

<DiscretizeBin binValue="0.92">

<Interval closure="openOpen" rightMargin="18.0"/>

</DiscretizeBin>

...

</Discretize>

</DerivedField>

<DerivedField dataType="double" displayName="Age" name="Age*" optype="continuous">

<Extension extender="KNIME" name="summary" value="Z-Score (Gaussian) normalization on 4 column(s)"/>

Page 157: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 157

157

<NormContinuous field="Age">

<LinearNorm norm="-2.8430412477523532" orig="0.0"/>

<LinearNorm norm="-2.769429284421807" orig="1.0"/>

</NormContinuous>

</DerivedField>

<DerivedField dataType="double" displayName="Income" name="Income*" optype="continuous">

<Extension extender="KNIME" name="summary" value="Z-Score (Gaussian) normalization on 4 column(s)"/>

<NormContinuous field="Income">

<LinearNorm norm="-1.2164099347279616" orig="0.0"/>

<LinearNorm norm="-1.2163955713786447" orig="1.0"/>

</NormContinuous>

</DerivedField>

<DerivedField dataType="double" displayName="Deductions" name="Deductions*" optype="continuous">

<Extension extender="KNIME" name="summary" value="Z-Score (Gaussian) normalization on 4 column(s)"/>

<NormContinuous field="Deductions">

<LinearNorm norm="-0.1983193433770224" orig="0.0"/>

<LinearNorm norm="-0.19538425035943" orig="1.0"/>

</NormContinuous>

</DerivedField>

<DerivedField dataType="double" displayName="Hours" name="Hours*" optype="continuous">

<Extension extender="KNIME" name="summary" value="Z-Score (Gaussian) normalization on 4 column(s)"/>

<NormContinuous field="Hours">

<LinearNorm norm="-3.297304369766391" orig="0.0"/>

<LinearNorm norm="-3.2150250058375485" orig="1.0"/>

</NormContinuous>

</DerivedField>

<DerivedField dataType="double" displayName="Age_binned" name="Age_binned*" optype="continuous">

<FieldRef field="Age_binned"/>

</DerivedField>

</LocalTransformations>

<PolynomialKernelType coef0="1.0" degree="1.0" gamma="1.0"/>

<VectorDictionary numberOfVectors="741">

<VectorFields numberOfFields="53">

<FieldRef field="Age*"/>

<FieldRef field="Income*"/>

Page 158: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 158

158

...

</VectorFields>

<VectorInstance id="1_1038288">

<REAL-SparseArray n="53">

<Indices>1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41

42 43 44 45 46 47 48 49 50 51 52 53</Indices>

<REAL-Entries>0.4694971021222236 -0.8179157566856157 -0.1983193433770224 1.228060646319936 1.0 0.0 0.0 0.0 0.0

0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

0.0 0.0 0.0 1.0 0.32</REAL-Entries>

</REAL-SparseArray>

</VectorInstance>

...

</VectorInstance>

...

</VectorDictionary>

<SupportVectorMachine targetCategory="0" alternateTargetCategory="1">

<SupportVectors numberOfAttributes="53" numberOfSupportVectors="741">

<SupportVector vectorId="1_1038288"/>

<SupportVector vectorId="1_1044221"/>

...

</SupportVectors>

<Coefficients numberOfCoefficients="741" absoluteValue="-1.9484983196017862">

<Coefficient value="1.0"/>

<Coefficient value="1.0"/>

...

</Coefficients>

</SupportVectorMachine>

</SupportVectorMachineModel>

Page 159: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 159

159

APPENDIX B – RESEARCH TOOLS

A collection of services and software aided the research during the construction of this

essay:

PRODUCTIVITY SOFTWARE

1. Microsoft Office 2013 Suite (Word, Excel and PowerPoint)

2. Microsoft Project 2013

3. Microsoft Visio 2013

INTERNET BROWSERS

4. Google Chrome internet browser

5. Internet Explorer internet browser

6. Firefox internet browser

OPEN SOURCE PREDICTIVE ANALYTICS AND DATA MINING TOOLS

7. R (The R Project for Statistical Computing, 2014)

8. Weka (Weka 3: Data Mining Software in Java, 2014)

9. PSPP (GNU PSPP, 2014)

10. Orange (Orange 2.7 for Windows, 2014)

11. KNIME (KNIME, 2014)

12. RapidMiner (RapidMiner STARTER Edition, 2014)

13. ELKI (LKI: Environment for Developing KDD-Applications Supported by Index-

Structures, 2014)

PYTHON RELATED STATISTICAL LIBRARIES

14. SciPy library Stack (SciPy Stack, 2014)

a. NumPy library (NumPy, 2014)

Page 160: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 160

160

b. pandas library (pandas, 2014)

c. SymPy library (SymPy, 2014)

d. IPython library (IPython, 2014)

15. scikit-learn library (scikit-learn: Machine Learning in Python, 2014)

16. scikit-image library (scikit-image: Image Processing in Python, 2014)

17. Matplotlib library (Matplotlib, 2014)

18. Statsmodels library (Statsmodels, 2014)

19. MpMath library (mpmath, 2014)

LITERATURE SEARCH ENGINES

20. Google (Google, 2014)

a. Google Correlate (Google Correlate, 2014)

b. Google Scholar (Google Scholar, 2014)

c. Google Trends (Google Trends, 2014)

21. WolfframAlpha (WolframAlpha, 2014)

22. Dogpile (dogpile, 2014)

23. iSeek at (iSeek Education, 2014)

24. refseek (refseek, 2014)

25. Virtual LRC (Virtual LRC, 2014)

26. AcademicIndex.net at (academicindex.net, 2014)

27. Digital Library of The Commons Repository (Digital Library of the Commons

Repository, 2014)

28. Microsoft Academic Research (Microsoft Academic Research, 2014)

Page 161: ATHABASCA UNIVERSITY A Survey of Predictive …dtpr.lib.athabascau.ca/action/download.php?filename=scis-07/open/... · ATHABASCA UNIVERSITY A Survey of Predictive Analytics in Data

A Survey of Predictive Analytics in Data Mining with Big Data 161

161

RESEARCH PAPER ONLINE DATABASES

29. IEEE Xplore Digital Library (IEEE Xplore Digital Library, 2014)

30. ACM Digital Library (ACM Digital Library, 2014)

31. EBSCO Colleges and Universities Online Resources for Academic Libraries

(EBSCO Colleges and Universities Online Resources for Academic Libraries,

2014)

32. IET Inspec Database (IET Inspec, 2014)

33. SpringerLink Database (SpringerLink, 2014)

34. Athabasca Online Library at (Athabasca University Library, 2014)

RESEARCH MANAGEMENT TOOLS AND SERVICES

35. Mendeley (Mendeley, 2014)

36. Qiqqa (Qiqqa, 2014)

37. Colwiz (Colwiz, 2014)

38. Evernote (Evernote, 2014)

39. Zotero (Zotero, 2014)

ONLINE COMMUNITIES

40. Academia.edu (Academia.edu, 2014)

41. Khan Academy (KhanAcademy Probability and Statistics, 2014)